# AI Insights for Humans

A working notebook of practical AI interpretability, prompting, reliability, and tooling notes for people who want evidence instead of hype.

Purpose: Practical AI insight notebook for human operators.
Entries: 67
Tags: agents, dictionary, evals, field, governance, integration, interpretability, notes, platform, prompting, reliability, security, tooling, verification

Public surfaces:
- HTML notebook: https://tips.platphormnews.com/
- Focus mode payload: https://tips.platphormnews.com/api/v1/focus-mode
- Dictionary/Notes integrations: https://tips.platphormnews.com/api/v1/integrations
- Platform literacy tips: https://tips.platphormnews.com/api/v1/platform-literacy
- Research trends and demos: https://tips.platphormnews.com/api/v1/research-trends
- Operator patterns: https://tips.platphormnews.com/api/v1/operator-patterns
- RSS: https://tips.platphormnews.com/rss.xml
- API docs: https://tips.platphormnews.com/api/docs
- OpenAPI: https://tips.platphormnews.com/openapi.yaml
- MCP: https://tips.platphormnews.com/api/mcp
- Full sitemap: https://tips.platphormnews.com/sitemap-full.xml
- Human attribution: https://tips.platphormnews.com/humans.txt

Protected actions require PLATPHORM_API_KEY.

## Route Standard
Implemented required routes: 22/22
- GET /: implemented; Public AI insights notebook and reading paths.
- GET /api/health: implemented; Platform health payload for the publication.
- GET /api/v1/health: implemented; Versioned platform health payload.
- GET /api/docs: implemented; Machine-readable API and route documentation.
- GET /api/v1/entries: implemented; Read-only JSON index of notebook entries.
- GET /api/v1/entries/{id}: implemented; Read-only JSON detail for one notebook entry.
- GET /api/v1/focus-mode: implemented; Read-only focus frames and practice loop payload.
- GET /api/v1/integrations: implemented; Read-only Dictionary and Notes integration map with public handoff boundaries.
- GET /api/v1/platform-literacy: implemented; Read-only interactive tips distilled from PlatPhormNews root discovery and policy surfaces.
- GET /api/v1/research-trends: implemented; Read-only source-backed AI research trends, public-safe demonstrations, and cross-model prompt contracts.
- GET /api/v1/operator-patterns: implemented; Read-only operator strengths, drag patterns, repair loops, and reusable workflow prompts.
- POST /api/v1/reports: implemented; Protected generation of a real notebook/platform summary report.
- GET /api/mcp: implemented; MCP metadata and JSON-RPC usage information.
- POST /api/mcp: implemented; JSON-RPC 2.0 MCP endpoint for real notebook tools, resources, and prompts.
- GET /openapi.yaml: implemented; OpenAPI 3.1 contract for public and protected routes.
- GET /llms.txt: implemented; Concise LLM-readable site summary.
- GET /llms-full.txt: implemented; Complete LLM-readable notebook and platform summary.
- GET /llms-index.json: implemented; Structured LLM index generated from real entries.
- GET /robots.txt: implemented; Crawler policy and sitemap pointer.
- GET /sitemap.xml: implemented; Canonical URL sitemap for the notebook.
- GET /sitemap-index.xml: implemented; Sitemap index for platform crawlers.
- GET /sitemap-full.xml: implemented; Complete sitemap covering product and machine-readable routes.
- GET /humans.txt: implemented; Human-readable ownership, purpose, attribution, and contact details.
- GET /rss.xml: implemented; RSS feed generated from notebook entries.
- GET /feed.xml: implemented; RSS-compatible feed alias.
- GET /atom.xml: implemented; Atom feed generated from notebook entries.
- GET /manifest.webmanifest: implemented; Web app manifest for install and preview surfaces.
- GET /.well-known/mcp.json: implemented; MCP capability discovery manifest.
- GET /.well-known/agents.json: implemented; Agent-facing route and policy discovery manifest.
- GET /.well-known/security.txt: implemented; Security contact and policy file.
- GET /.well-known/trust.json: implemented; Trust, auth, public/protected, and exposure policy.

## Trust Policy
Web dashboard, public-safe discovery, browser-based operations, trusted-domain discovery, standard route compliance, Vercel metadata capture, trace inspection, and agentic workflow discovery are intentionally supported for public read-only debugging and operator workflows. Mutating, administrative, ingestion, replay, fork, remediation, deployment, sync, test-triggering, reporting, and write actions require PLATPHORM_API_KEY.

## Field Guide Playbooks
### Hallucination triage
A model produced a confident answer and nobody knows whether the weak point is retrieval, reasoning, or source support.
Outcome: A reviewer can identify which claims are supported, which claims need receipts, and which answer parts should be withheld.
1. Locate fragile spans: Hallucinations often hide where fluency is highest but token certainty is thinnest.
2. Separate retrieval from generation: Bad context can make a good model sound wrong with confidence.
3. Attach receipts: The unsupported claim does not vanish; it becomes visible enough to reject.
4. Decide the confidence bucket: Confidence matters when it changes the next action.

### Prompt change control
A system prompt needs improvement, but every wording edit risks changing refusal, format, or evidence behavior.
Outcome: The prompt change ships with a semantic diff, targeted fixtures, and a clear before/during/after eval record.
1. Shrink the contract: A smaller system prompt is not weaker; it is easier to audit.
2. Diff the meaning: A one-word modal change can quietly become a policy change.
3. Run the eval sandwich: The middle eval catches drift while the edit is still cheap.
4. Mutate the boundary: A boundary that only catches the obvious phrasing is not a boundary.

### Agent action gate
An agent can call tools, but the operator needs confidence before it mutates, publishes, replays, deletes, or spends.
Outcome: Every action moves through preview, inspect, authorize, execute, and verify without hiding protected states.
1. Preview the action: Permission to plan is not permission to execute; the preview is where unsafe scope becomes inspectable.
2. Respect blocked states: A designed stop is better than a graceful guess because it preserves trust and gives the operator a next move.
3. Spend checks by risk: Latency can be evidence-driven instead of uniformly expensive.
4. Verify after execution: The world, not the model, gets the final vote, so post-action evidence is part of the action itself.

## Failure Case Studies
### The source-shaped hallucination
Setup: A model summarizes a report and includes a link beside every claim.
Failure: The links exist, but two citations point to documents that discuss the topic without supporting the exact sentence.
Hidden signal: The unsupported claims are entity/date claims with weak token margins and broad source references.
Repair: Use receipts per claim, then triangulate the high-impact facts with independent sources before publishing.
Aha: A citation beside a sentence is not the same thing as support under the sentence.

### The buried constraint
Setup: A long prompt includes a critical policy exception in the middle of retrieved context.
Failure: The answer follows the opening task and closing format while ignoring the exception.
Hidden signal: The missed fact was technically present but surrounded by lower-value context.
Repair: Move the task contract to the opening, repeat the exception at the end, and rank retrieved chunks before generation.
Aha: Presence in context is not the same thing as usable salience.

### The tool call that did too much
Setup: An agent is asked to clean a set of records and calls a bulk mutation tool.
Failure: The selector is broader than the user intended and changes records outside the target set.
Hidden signal: The model never previewed target ids or stated a rollback condition.
Repair: Add a dry-run gate, require explicit target ids, and make protected-required a first-class state.
Aha: Tool safety starts before execution, not after the audit log catches damage.

### The helpful prompt edit that moved the boundary
Setup: A prompt is edited to make answers friendlier and less formal.
Failure: The model starts complying with requests it previously redirected because the new examples imply looser policy.
Hidden signal: The text diff is small, but the semantic diff changes tone, authority, and refusal posture.
Repair: Run a semantic diff, then use refusal fixtures and harmless mutants before shipping.
Aha: Style examples can become behavioral instructions if you do not test the boundary.

## AHA Ladder
### From answer to signal
Before: Read the model's sentence and judge whether it sounds plausible.
After: Inspect token margins, retrieved evidence, receipts, and contradiction before judging.
Operator question: Where would this answer break first?

### From prompt to contract
Before: Write instructions until the behavior feels described.
After: Define the output schema, failure states, hard boundary, and test fixtures.
Operator question: What is the model allowed to leave blank?

### From tool to action
Before: Let the agent call a tool when it appears to understand the task.
After: Preview arguments, inspect target ids, authorize mutation, execute, then verify changed state.
Operator question: What would stop this action before it runs?

### From confidence to decision
Before: Ask the model for a percentage and decide whether it feels high enough.
After: Bucket confidence by evidence quality, conflict level, and downstream impact.
Operator question: What does this confidence label change?

## Focus Mode
### Name the term before the trick
Takeaway: Model behavior is hard to inspect when the key vocabulary is unstable.
Dictionary term: retrieval relevance
Notes prompt: Which word in today's demo felt familiar but turned out to need a stricter definition?
Aha: The glossary is not background material. It is part of the operating system for the work.

### Put an artifact on screen
Takeaway: The room needs something concrete to inspect: an answer, citation, tool preview, or note.
Dictionary term: receipt
Notes prompt: What exact artifact made you change your mind about the model output?
Aha: Evidence has to be visible before judgment can become repeatable.

### Preview action before authority
Takeaway: A model that can mutate state needs a dry run, target ids, stop conditions, and verification.
Dictionary term: dry run
Notes prompt: What would stop this tool call before it touches real state?
Aha: Permission to explain a tool call is not permission to execute it.

### Make stopping a valid answer
Takeaway: Unknown, partial, protected-required, and needs-review states keep bad certainty from becoming product behavior.
Dictionary term: abstention
Notes prompt: Where did the model need an honest stop state instead of a smoother sentence?
Aha: A graceful stop can be the most helpful output in the room.

### Turn repeated confusion into better tools
Takeaway: Repeated confusion is a product signal, not an interruption.
Dictionary term: provenance
Notes prompt: What phrase should become a shared glossary candidate after this walkthrough?
Aha: The field notes are generating the next version of the shared vocabulary.

### End with recall, artifact, transfer
Takeaway: An insight is useful when it can be named, used on an artifact, and transferred to another workflow.
Dictionary term: prompt fixture
Notes prompt: Write one transfer question that would catch shallow memorization.
Aha: The final question should prove the habit moved beyond this frame.

## Dictionary and Notes Integrations
### PlatPhorm Dictionary
Public terminology, definitions, provenance lookup, search, datasets, term graph, and MCP discovery.
Public reads: Approved public terms, Dictionary search, OpenAPI and llms discovery, Term graph and provenance lookup, Read-only MCP introspection
Protected boundary: Moderation, import, admin, private submissions, audit logs, protected reports, model-assisted moderation, registry mutation, and destructive actions require PLATPHORM_API_KEY when enforcement is enabled.
Notebook use: This notebook links concepts to Dictionary as concept receipts and glossary anchors; it does not publish definitions on the user's behalf.

### PlatPhorm Notes
Public learning notepad for note enrichment, terminology detection, Dictionary lookup, and human-approved definition staging.
Public reads: App shell and discovery files, Public templates, Term detection previews, Note analysis previews, Workflow templates and integration status
Protected boundary: No visitor key entry. Backend calls to Dictionary, Docs, OpenContent, Sheets, Phorm, SVG, and other PlatPhorm services require PLATPHORM_API_KEY server-side.
Notebook use: This notebook gives readers prompts and observation formats to use in Notes; it does not store note bodies or stage submissions directly.

## PlatPhormNews Root Interactive Tips
### Public DMZ means safe read-only front door
Surface: Root MCP manifest and trust policy
Source: https://platphormnews.com/.well-known/mcp.json
Plain language: Think of the public DMZ like a public lobby. Anyone can ask where things are, but they cannot unlock private rooms, edit records, or trigger protected work.
Operator: The root site publishes public discovery, network graph, stories, route compliance, and MCP introspection while reserving sync, validation, reports, and policy updates for protected actions.
Real-world view: Your job is to check that the front desk gives useful directions without accidentally becoming an admin panel.
Aha: Public does not mean powerful. Public means safely inspectable.
Safe boundary: Do not test destructive, private, protected, credential, admin, or mutation paths without explicit authorization.

### The network graph is the map of the city
Surface: GET /api/network/graph
Source: https://platphormnews.com/api/network/graph
Plain language: A website network is like a city. The graph tells you what buildings exist and how roads connect them.
Operator: The root graph reports sites, nodes, edges, sources, categories, route compliance, discovery compliance, and trusted domain status.
Real-world view: A graph helps you check coverage: which sites are trusted, which surfaces are discoverable, and which claims need source-backed evidence.
Aha: Discovery is not guessing URLs. Discovery is reading the map the platform publishes.
Safe boundary: Stay inside trusted domains and published discovery routes. Do not crawl arbitrary user-provided URLs.

### robots.txt and llms files teach crawler manners
Surface: robots.txt, llms.txt, llms-full.txt, llms-index.json
Source: https://platphormnews.com/robots.txt
Plain language: robots.txt is a hallway sign for crawlers. llms files are a study guide for AI systems.
Operator: The root robots policy allows public discovery files and graph routes for LLM crawlers while blocking admin, private, protected, auth, cron, webhook, and secret-query paths.
Real-world view: Good automation respects the published crawl lanes. Ignoring them is not clever; it is unsafe and low-signal.
Aha: The safest crawler is the one that reads instructions before touching the site.
Safe boundary: Never put tokens, keys, or secrets into URLs, traces, screenshots, feeds, or public artifacts.

### MCP is a toolbox, not a permission slip
Surface: .well-known/mcp.json and /api/mcp
Source: https://platphormnews.com/.well-known/mcp.json
Plain language: MCP lists tools an AI client can ask for, like a toolbox label. The label does not mean every tool is safe for every person to use.
Operator: The root manifest exposes JSON-RPC methods, public tools, protected tools, resources, prompts, trace headers, and the PLATPHORM_API_KEY boundary.
Real-world view: Read tools/list and resources/list first. Confirm whether a tool is public read-only or protected before designing any workflow.
Aha: Tool discovery comes before tool execution.
Safe boundary: Do not call sync, refresh, ingestion, reporting, or policy-update tools unless authorized and key-gated.

### OpenAPI is the menu with ingredients listed
Surface: /api/docs and /openapi.yaml
Source: https://platphormnews.com/api/docs
Plain language: An API doc is a menu. It tells you what you can ask for, what parameters it accepts, and what shape the answer should have.
Operator: The root API documentation covers stories, search, sources, network APIs, network MCP, evidence, trustops, searchops, agentops, public contract, feeds, and discovery routes.
Real-world view: A contract lets you test behavior without guessing. You can verify route shape, auth requirements, and public-safe responses.
Aha: Guessing endpoint behavior is amateur hour; reading the contract is how experts move fast safely.
Safe boundary: Use read-only endpoints for practice. Treat POST, sync, ingestion, and reporting routes as protected unless explicitly documented otherwise.

### JSON-LD is the site's ID card
Surface: Homepage JSON-LD
Source: https://platphormnews.com
Plain language: JSON-LD is structured data that tells search engines and agents who the site is, what it publishes, and what actions it supports.
Operator: The root homepage publishes organization, website search action, navigation, and FAQ structured data for public identity and discoverability.
Real-world view: Structured data is a claim surface. Compare it with live routes and policies to catch stale identity, stale navigation, or unsafe promises.
Aha: A page can speak to humans in HTML and to machines in JSON-LD at the same time.
Safe boundary: Do not treat structured data as secret or authority by itself; verify with the corresponding public route.

### RSS tells what changed; sitemaps tell what exists
Surface: RSS, Atom, sitemap, sitemap index
Source: https://platphormnews.com/rss.xml
Plain language: RSS is the update stream. A sitemap is the table of contents. Together they help people and machines find new and existing pages.
Operator: The root publishes RSS/feed/Atom and several sitemap variants, including network and full sitemap surfaces, with robots pointers.
Real-world view: Feeds and sitemaps are low-risk coverage tools. They reveal intended public pages without probing private paths.
Aha: The public site tells you where to look if you start with the sitemap instead of guessing.
Safe boundary: Use published sitemaps instead of brute-force path discovery.

### Agent policy tells bots how to behave
Surface: .well-known/agent-policy.json
Source: https://platphormnews.com/.well-known/agent-policy.json
Plain language: Agent policy is the operating boundary for AI clients: what they can read, what needs a key, and what data must stay private.
Operator: The root policy names public access surfaces, supported client classes, PLATPHORM_API_KEY requirements, trace propagation, JA4 redaction, and no-private-data rules.
Real-world view: Before running any automated inspection, prove your agent obeys the policy and can stop at protected boundaries.
Aha: A well-behaved agent is safer because it knows where not to go.
Safe boundary: No secrets, cookies, raw IP addresses, raw JA4 digests, private traces, or private audit details in public reports.

### Trace headers are the breadcrumb trail
Surface: MCP manifest and agent policy trace fields
Source: https://platphormnews.com/.well-known/agent-policy.json
Plain language: Trace headers are like numbered sticky notes that help you follow one request as it moves between services.
Operator: Root policy accepts traceparent, tracestate, X-PlatPhorm-Trace-Id, and propagates safe PlatPhorm trace/request/source/target headers.
Real-world view: Traceability helps debug cross-site behavior, but secret baggage must never ride along.
Aha: Observability is useful only when it does not leak the thing it observes.
Safe boundary: Never put authorization, API keys, cookies, session tokens, or raw private payloads into trace state or baggage.

### Status should say how it knows
Surface: Network graph and compliance payloads
Source: https://platphormnews.com/api/network/graph
Plain language: A status card is better when it tells you where the number came from, not just whether it is green.
Operator: Root discovery surfaces label sources such as static manifest, MCP registry, sitemap index, declared route spec, stale markers, confidence, and checkedAt timestamps.
Real-world view: Source labels let you separate current evidence from stale, degraded, inferred, or declared-only status.
Aha: Green without evidence is decoration. Green with source and time is operational signal.
Safe boundary: Do not convert degraded or declared-only status into a claim of live success.

## Skill Ramp
### First-pass reader
Reader can: Explain the difference between public read, protected action, and private data.
Practice: Sort routes from robots.txt into safe-to-read and stop-sign groups.

### Careful operator
Reader can: Use sitemaps, RSS, llms files, OpenAPI, and JSON-LD to inspect a site without guessing.
Practice: Build a source-backed checklist for one public surface.

### Agent workflow designer
Reader can: Read MCP tools/resources/prompts and design a public-safe workflow with explicit stop states.
Practice: Write a dry-run plan for a read-only MCP inspection.

### Real-world operator
Reader can: Audit public automation behavior while respecting trust policy, robots policy, auth boundaries, trace redaction, and no-secret rules.
Practice: Produce a report that says what was inspected, what was skipped, why it was skipped, and what evidence supports each conclusion.

## Research Trends and Demonstrations
### Seeing inside AI means evidence, not telepathy
Plain language: We do not get the model's private thoughts. We get clues: token probabilities, attention paths, features, circuits, traces, and tool receipts.
Advanced signal: Attribution graphs, activation patching, logit lenses, and sparse features are turning interpretability into causal hypothesis testing.
Aha: The mind of AI is most visible when you compare what changed inside the model with what changed in the answer.
Operator move: Treat every visualization as a hypothesis until an intervention or independent check changes the behavior.
Sources: anthropic-circuit-tracing, anthropic-attribution-graphs

### Loops beat one-shot answers for serious work
Plain language: A stronger AI workflow observes, decides, acts, checks, and revises. The check is where the magic becomes engineering.
Advanced signal: ReAct-style agent loops, evaluator-optimizer workflows, and long-task evals all point to sustained control loops rather than single completions.
Aha: A loop is not repetition. It is a chance to compare the model's plan with the world before the next move.
Operator move: Design stop conditions, retry limits, and evidence checks before granting tools or longer task horizons.
Sources: react-paper, anthropic-effective-agents, metr-long-tasks

### Context engineering is the new prompt engineering
Plain language: The prompt is more than the words you type. It is the whole working set the model sees: files, memories, tools, examples, and rules.
Advanced signal: Agent quality is increasingly constrained by context curation, source ordering, compression, and boundary placement.
Aha: If the right fact is buried in the wrong place, the model can technically have it and still act like it never saw it.
Operator move: Build a source map before summary and put the highest-risk constraints near the edges of context.
Sources: anthropic-context-engineering, simon-context-engineering

### Agents need protocols before broad authority
Plain language: Before a model acts, it needs to know what tools exist, what schema they require, and what it is allowed to do.
Advanced signal: MCP, A2A, managed agent tools, and trace surfaces are converging on explicit contracts around tools and agent-to-agent work.
Aha: Tool discovery is a safety feature because it turns invisible authority into inspectable capability.
Operator move: Require schema inspection and protected boundary checks before execution.
Sources: anthropic-mcp, google-a2a, openai-agents-tools

### OpenClaw is best taught as a visible toolchain
Plain language: OpenClaw is understandable when the chain is visible: discover, inspect, validate, dry-run, protect, then execute only when authorized.
Advanced signal: The PlatPhorm OpenClaw gateway should expose public-safe metadata and dry-run plans while keeping mutations behind PLATPHORM_API_KEY.
Aha: The valuable question is not did the agent act. It is whether the chain proved the agent should be allowed to act.
Operator move: Make dry-run evidence the artifact reviewed before any workflow or remediation action.
Sources: openclaw-platform, anthropic-mcp

### Webhooks teach cause, delivery, and proof
Plain language: A webhook is a message from one system to another. The signature and event contract tell you whether to trust it.
Advanced signal: Webhook reliability depends on signing, idempotency, event schemas, delivery logs, retries, and replay boundaries.
Aha: The event is not the proof. The verified signature, contract match, delivery receipt, and replay policy are the proof.
Operator move: Use public-safe signature and contract checks for practice; require PLATPHORM_API_KEY for event creation, replay, or mutation.
Sources: webhooklab-platform, owasp-llm-top-10

### Prompts that travel are contracts, not slogans
Plain language: OpenAI, Claude, and Gemini differ, but the best prompts give all of them the same useful structure: role, task, context, constraints, evidence, output shape, and stop rules.
Advanced signal: Official provider guidance keeps converging on explicit instructions, examples, structured outputs, testing, and clear context boundaries.
Aha: A cross-model prompt works when the model can tell what counts as complete, partial, unknown, and unsafe.
Operator move: Version prompt contracts and test them against local fixtures instead of rewriting them for model brand voice.
Sources: openai-prompt-guide, anthropic-prompt-guide, gemini-prompt-guide

### Reasoning models make search visible, not automatically true
Plain language: Reasoning models can spend more time exploring. That is useful, but the final answer still needs receipts.
Advanced signal: Reinforcement learning for reasoning and test-time compute are changing how models allocate effort, especially on math, code, and multi-step tasks.
Aha: A longer reasoning path is not proof. Proof is the artifact that survives checking.
Operator move: Escalate to reasoning models where uncertainty, task horizon, or decision impact justifies the extra latency.
Sources: deepseek-r1, swe-bench-verified

### The new evals are failure maps
Plain language: Benchmarks are useful when they show where models still fail, not when they become scoreboard decoration.
Advanced signal: Humanity's Last Exam, SWE-bench Verified, and long-task evaluations push toward harder, more realistic failure discovery.
Aha: A benchmark score is less useful than the broken examples it helps you test in your own product.
Operator move: Translate public benchmark failure patterns into small product-specific fixtures.
Sources: humanitys-last-exam, swe-bench-verified, metr-long-tasks

### AI security lives at the text-tool boundary
Plain language: The risky moment is often when text becomes action: a prompt changes a tool call, a source injects instructions, or an output gets trusted too quickly.
Advanced signal: Prompt injection, excessive agency, insecure output handling, and data leakage are now baseline LLM app risks.
Aha: The model is safest when it can read broadly, plan visibly, act narrowly, and stop cleanly.
Operator move: Bind tools to trusted domains, schema validation, least authority, and protected action checks.
Sources: owasp-llm-top-10, openai-agents-tools

## Research Demo Tracks
### Loops: Observe, decide, act, check, reflect
What changes when AI stops answering once and starts checking itself against the world?
First-pass model: A loop is like solving a math problem, checking the work, and using the mistake to choose the next step.
Operator model: A loop is a bounded controller with state, observations, tool actions, verifier signals, stop rules, and retry budgets.
Aha: The check step is where AI becomes inspectable. Without it, the loop is just a confident answer repeated several times.
Boundary: This demo explains loop mechanics and public-safe dry runs. It does not trigger external tools or protected actions.
1. Observe: Read the question, sources, tool state, and current constraints. Signal: Source map, current state, missing evidence.
2. Decide: Choose one next move that can be verified. Signal: Plan line with target, reason, and stop condition.
3. Act: Run a read, draft, calculation, or protected tool only inside the boundary. Signal: Tool call preview or public-safe artifact.
4. Check: Compare the result with expected evidence. Signal: Pass, mismatch, missing receipt, or retry reason.
5. Reflect: Update the next step, confidence, or refusal branch. Signal: Changed plan, not just longer prose.

### OpenClaw: Capability discovery before agent authority
How do we show an agent workflow without pretending execution happened?
First-pass model: OpenClaw is the checklist a tool-using agent walks through before it is allowed to touch the real system.
Operator model: The PlatPhorm OpenClaw surface is an OpenClaw-compatible planning and validation gateway, not a claim that every hosted runtime mode is live.
Aha: OpenClaw is easiest to understand as a chain of evidence. Each step earns the right to ask for the next one.
Boundary: This site only demonstrates the OpenClaw-compatible flow and links to the platform surface. Real execution belongs behind PLATPHORM_API_KEY on the owning service.
1. Discover: Find the service purpose, MCP/API metadata, and public capabilities. Signal: Capability list and service boundary.
2. Inspect: Read tool schemas, required inputs, and protected fields. Signal: Arguments, allowed values, and refusal states.
3. Draft: Create a toolchain plan with target ids and expected output. Signal: Dry-run plan, not a success claim.
4. Validate: Check schema fit, trusted domain, and route compliance. Signal: Pass, degraded, or protected_required.
5. Handoff: Prepare protected execution only when authorized. Signal: PLATPHORM_API_KEY boundary and verification artifact.

### Webhooks: Event, signature, delivery, retry, replay
What proof should exist before a receiving system trusts an event?
First-pass model: A webhook is a doorbell from another app. The signature proves who rang it, and the event shape proves what they meant.
Operator model: A production webhook contract includes schema validation, signing, idempotency, delivery logs, retry policy, replay rules, and redacted artifacts.
Aha: The receiving system should trust the verified contract, not the model's summary of the event.
Boundary: This notebook demonstrates lifecycle thinking. Signature generation, event creation, replay, and delivery live on WebhookLab with its own public/protected boundary.
1. Contract: Name event type, schema, source, and receiver expectations. Signal: Event contract and required fields.
2. Sign: Generate or verify an HMAC-style signature with a secret kept out of public surfaces. Signal: Signature status, never raw secret.
3. Deliver: Send to receiver and record delivery metadata. Signal: Status code, timestamp, redacted headers.
4. Verify: Check signature, event id, timestamp tolerance, and schema. Signal: Accepted, rejected, duplicate, or stale.
5. Replay: Retry or replay only through the owning protected service. Signal: Replay request id and reason.

### Premier model prompts: One contract, many model families
What kind of prompt survives OpenAI, Claude, Gemini, and future frontier models?
First-pass model: A good prompt is a worksheet: the model knows the job, the sources, the rules, the answer boxes, and when to say unknown.
Operator model: Portable prompts separate task contract from provider-specific affordances: role, objective, context ledger, constraints, output schema, uncertainty, and stop conditions.
Aha: The portable prompt is not the prettiest wording. It is the clearest contract for evidence, limits, and output shape.
Boundary: This demo provides prompt templates. It does not call external model APIs from the public page.
1. Role: Name the useful role and the decision the answer supports. Signal: The model knows who the answer is for.
2. Evidence: Provide allowed sources and what each can prove. Signal: Source ledger before final prose.
3. Constraints: State boundaries, exclusions, and protected actions. Signal: Clear public/protected/private split.
4. Schema: Define fields, allowed states, and citation shape. Signal: The answer can be tested mechanically.
5. Review: Ask for counter-evidence, confidence reason, and next check. Signal: The answer exposes its weakest point.

### AI mind microscope: Signals, circuits, traces, receipts
What can we honestly see when we say we want to look inside the model?
First-pass model: Imagine a dark room with instruments. You cannot see everything, but each instrument shows one kind of signal.
Operator model: Interpretability combines output-side telemetry, internal activation evidence, causal interventions, and external tool traces.
Aha: Seeing inside AI means layering instruments. No single view is the mind; together they make behavior less mysterious.
Boundary: This demo is explanatory. It does not expose private chain-of-thought, hidden provider internals, or live model traces.
1. Tokens: Inspect low-margin tokens and alternatives. Signal: Where the answer was fragile.
2. Attention: Watch which tokens route information to which positions. Signal: Possible routing paths.
3. Features: Probe or inspect sparse features associated with a concept. Signal: Concept directions and activation examples.
4. Circuits: Trace and patch components to test causal contribution. Signal: Behavior changes after intervention.
5. Receipts: Check external sources, tests, and tool traces. Signal: Evidence that survives outside the model.

## Cross-Model Prompt Patterns
### Evidence-first answer
Works across: OpenAI GPT/reasoning models, Anthropic Claude, Google Gemini
Use when: Research, factual analysis, route audits, policy summaries, and any answer where source support matters.
Why it works: It makes source authority explicit before prose, which reduces the model's incentive to blend evidence with memory.
Failure mode: The model may cite adjacent background unless you require claim-level receipts.
Template:
- Task: Answer the question for [audience] using only the sources below.
- Source ledger: For each source, state what it can prove and what it cannot prove.
- Answer fields: answer, supporting receipts, counter-evidence, uncertainty, next check.
- If evidence is missing, return status: insufficient_evidence instead of guessing.

### Tool dry-run contract
Works across: OpenAI agents, Claude tool use, Gemini function calling
Use when: Any workflow that could mutate, send, replay, publish, delete, spend, or trigger another system.
Why it works: It separates permission to plan from permission to execute, which is the core tool-safety boundary.
Failure mode: A model may treat planning approval as execution approval unless the stop condition is explicit.
Template:
- Before any tool call, produce a dry_run object.
- Include tool_name, target_ids, inputs, expected_effect, protected_boundary, stop_conditions, verification_plan.
- Do not execute if target_ids are broad, secrets are present, or protected auth is missing.
- Return protected_required when the action requires PLATPHORM_API_KEY.

### Cross-model comparison
Works across: OpenAI, Anthropic, Google, open-weight reasoning models
Use when: Prompt benching, model routing, field demos, and high-impact decisions where model disagreement is useful.
Why it works: The same contract makes failures comparable and turns disagreement into a source of insight.
Failure mode: Changing the prompt between models invalidates the comparison.
Template:
- Run the same task contract on each model without changing source material.
- Score outputs by fields: follows_schema, uses_receipts, names_uncertainty, avoids_overreach, helpful_next_action.
- Report disagreements before selecting a winner.
- Route by observed failure mode, not brand preference.

### Long-context source map
Works across: Long-context GPT models, Claude, Gemini
Use when: Large documents, repo audits, policy packs, discovery files, and multi-source research.
Why it works: It counters context rot by ranking evidence and making unsupported spans easy to spot.
Failure mode: A giant context window can still bury decisive facts in the middle.
Template:
- First build a source map with source_id, owner, freshness, what_it_proves, limitations.
- Then answer using source_id references next to each material claim.
- Put critical constraints at the opening and repeat the highest-risk one at the end.
- Do not summarize sources that were not inspected.

### AHA check
Works across: ChatGPT, Claude, Gemini, local assistants
Use when: Framing concepts that need to ramp from first exposure to operator-level judgment.
Why it works: It forces the model to teach mechanism, artifact, failure, and transfer instead of producing a polished mini-essay.
Failure mode: The explanation may stay inspirational unless the inspected artifact is specific.
Template:
- Explain the concept in plain language in three sentences.
- Show the concrete artifact the reader should inspect.
- Name one safe mistake and the repair habit.
- End with: Before I thought ___. Now I inspect ___. Next I will ___.

## Research Source Ledger
### Tracing the thoughts of a large language model
Publisher: Anthropic
Kind: lab_blog
URL: https://www.anthropic.com/news/tracing-thoughts-language-model
Signal: Attribution graphs and circuit tracing are turning model internals into inspectable pathways rather than mystery boxes.
Reader takeaway: We cannot read private thoughts, but we can inspect partial evidence about which features and paths influenced an answer.

### Attribution graph methods
Publisher: Transformer Circuits / Anthropic
Kind: paper
URL: https://transformer-circuits.pub/2025/attribution-graphs/methods.html
Signal: Mechanistic interpretability is becoming a reproducible workflow: build graph, localize path, test causal contribution.
Reader takeaway: A beautiful graph is only a hypothesis until an intervention changes model behavior.

### ReAct: Synergizing reasoning and acting in language models
Publisher: arXiv
Kind: paper
URL: https://arxiv.org/abs/2210.03629
Signal: The reason-act-observe rhythm remains the clearest first-pass model for agent loops.
Reader takeaway: An agent is not magic autonomy. It is a loop with observations, decisions, actions, and checks.

### Building effective agents
Publisher: Anthropic Engineering
Kind: lab_blog
URL: https://www.anthropic.com/engineering/building-effective-agents
Signal: Useful agents are usually simple workflows with clear tool boundaries before they are open-ended autonomy.
Reader takeaway: Start with a visible workflow, then let the model choose inside well-labeled rails.

### Effective context engineering for AI agents
Publisher: Anthropic Engineering
Kind: lab_blog
URL: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Signal: The practical prompt surface is the whole context stack: instructions, tools, memory, retrieved files, and stop rules.
Reader takeaway: Prompting is not one clever sentence. It is arranging the evidence and boundaries the model can see.

### Context engineering
Publisher: Simon Willison
Kind: industry_blog
URL: https://simonwillison.net/2025/Jun/27/context-engineering/
Signal: Practitioners are naming context engineering as the craft of assembling the right prompt-time working set.
Reader takeaway: If the model ignores a fact, ask whether the fact was positioned, formatted, and prioritized like it mattered.

### Introducing the Model Context Protocol
Publisher: Anthropic
Kind: protocol
URL: https://www.anthropic.com/news/model-context-protocol
Signal: Model-to-tool interoperability is shifting from one-off connectors toward explicit protocols.
Reader takeaway: Before a model uses a tool, it should discover what the tool is, what schema it accepts, and what authority it has.

### A2A: A new era of agent interoperability
Publisher: Google Developers Blog
Kind: protocol
URL: https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
Signal: Agent interoperability is becoming a platform concern, not only a prompt-design concern.
Reader takeaway: When two agents cooperate, the contract between them matters as much as either model.

### New tools for building agents
Publisher: OpenAI
Kind: official_docs
URL: https://openai.com/index/new-tools-for-building-agents/
Signal: Agent products are consolidating around model responses, tool calls, tracing, and managed action surfaces.
Reader takeaway: A serious agent demo should show the tool call, the trace or receipt, and the boundary, not only the final answer.

### Prompt engineering guide
Publisher: OpenAI Platform
Kind: official_docs
URL: https://platform.openai.com/docs/guides/prompt-engineering
Signal: Official prompt guidance keeps converging on clear instructions, context, examples, and explicit output contracts.
Reader takeaway: The prompt that travels across models is the one that names the job, evidence, limits, and output shape.

### Prompt engineering overview
Publisher: Anthropic Docs
Kind: official_docs
URL: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
Signal: Provider-specific details differ, but task clarity, examples, role, and format remain stable cross-model levers.
Reader takeaway: Do not write a prompt for one brand personality. Write a contract a capable model can follow.

### Gemini prompting strategies
Publisher: Google AI for Developers
Kind: official_docs
URL: https://ai.google.dev/gemini-api/docs/prompting-strategies
Signal: Gemini guidance reinforces the same core pattern: clear task, sufficient context, examples, constraints, and testing.
Reader takeaway: Cross-model prompts work best when they are boringly explicit about success and failure.

### DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning
Publisher: arXiv
Kind: paper
URL: https://arxiv.org/abs/2501.12948
Signal: Reasoning models made test-time search, verification, and longer deliberation mainstream product knobs.
Reader takeaway: The model may spend extra compute exploring possible paths. The user still needs evidence for the final path.

### Humanity's Last Exam
Publisher: arXiv
Kind: paper
URL: https://arxiv.org/abs/2501.14249
Signal: Frontier evals are becoming stress tests for knowledge, reasoning, calibration, and overconfidence.
Reader takeaway: A hard benchmark is not a product guarantee. It is a map of where models still break.

### Measuring AI ability to complete long tasks
Publisher: METR / arXiv
Kind: paper
URL: https://arxiv.org/abs/2503.14499
Signal: Agent progress is increasingly measured by task horizon, recovery from errors, and sustained work over time.
Reader takeaway: The question is not only can the model answer. It is how long it can keep a goal alive without drifting.

### Introducing SWE-bench Verified
Publisher: OpenAI
Kind: paper
URL: https://openai.com/index/introducing-swe-bench-verified/
Signal: Coding-agent evaluation is moving toward reviewed, reproducible tasks with real repositories and tests.
Reader takeaway: A coding answer is not done when it sounds right. It is done when the repo evidence turns green.

### OWASP Top 10 for Large Language Model Applications
Publisher: OWASP
Kind: security_standard
URL: https://owasp.org/www-project-top-10-for-large-language-model-applications/
Signal: Prompt injection, insecure output handling, and excessive agency are now standard LLM application risks.
Reader takeaway: Security starts at the boundary between text, tools, data, and authority.

### OpenClaw-compatible PlatPhorm gateway
Publisher: PlatPhormNews
Kind: platform_surface
URL: https://claws.platphormnews.com
Signal: OpenClaw is useful here as an integration lens: discover capabilities, draft toolchains, dry-run, validate, then protect execution.
Reader takeaway: The valuable demo is not pretending a tool ran. It is showing the chain that decides whether it should run.

### WebhookLab public-safe contract testing
Publisher: PlatPhormNews
Kind: platform_surface
URL: https://webhooklab.platphormnews.com
Signal: Webhook demos can teach signing, delivery, verification, retry, and replay without sending unsafe live events.
Reader takeaway: A webhook is an agreement plus a receipt. If you cannot verify the signature and event shape, do not trust the delivery.

## Operator Patterns: What To Copy
### Raise the bar after the first green check
A first working version is treated as the start of the real pass, not the finish line.
Looks like: After build or deploy succeeds, the next request pushes into content depth, routes, discovery, tests, mobile behavior, and live surfaces.
Why it works: Most weak AI products fail in the gap between local success and operator-grade usefulness. This habit keeps pressure on the gap.
Failure mode: The finish line can move faster than the team can name what done means.
Repair move: Convert the next push into a named acceptance target: content, route, test, live smoke, or cross-site handoff.
Practice move: Take a green build and list the evidence still missing before calling the product ready.

### Reject fake success aggressively
If a system did not send, run, publish, replay, or verify something, the UI should not pretend it did.
Looks like: Buttons, reports, demo tracks, MCP tools, and status cards need to say public-safe, protected_required, degraded, unsupported, or ready with real evidence.
Why it works: AI systems get dangerous when fluent summaries hide missing execution. Honest state labels preserve trust.
Failure mode: The product can feel less magical because it refuses to fake the last mile.
Repair move: Make the boundary useful: show the dry run, the missing auth, the required receipt, and the next real action.
Practice move: Take five shiny UI claims and label each one real, simulated, protected, degraded, or fake.

### Think in platform contracts
Every site still does its own job, but it also needs standard discovery, health, auth, docs, feeds, and machine-readable surfaces.
Looks like: A feature request quickly turns into questions about OpenAPI, llms files, MCP, RSS, sitemap, trust policy, route compliance, and tests.
Why it works: Humans get the usable page. Agents get the contract. Operators get a way to verify both.
Failure mode: The shared platform layer can crowd out the site's unique product if the purpose is not named first.
Repair move: Start each pass with the site role, then add only the platform surfaces that make that role more usable or inspectable.
Practice move: Pick a small app and write its unique purpose before listing the standard platform routes it should expose.

### Think in networks, not isolated pages
A useful AI learning surface should connect to definitions, notes, workflows, traces, evals, webhooks, and source ledgers where they actually help.
Looks like: The same concept gets linked to Dictionary, Notes, MCP, WebhookLab, OpenClaw, BrowserOps, Evals, and public JSON when those handoffs are real.
Why it works: AI work is rarely a single-page event. Learning, evidence, action, and verification usually live in different systems.
Failure mode: Too many integrations can become decorative navigation instead of real workflow support.
Repair move: Every integration card must name owner, public reads, protected actions, and what this page does not do.
Practice move: Map one AI investigation across term definition, note capture, tool dry run, route check, and evidence report.

### Build the ramp from plain reading to operator judgment
The content should start in plain language and still scale toward advanced AI system judgment.
Looks like: A walkthrough starts with plain terms, then adds artifacts, receipts, schemas, boundary checks, evals, and cross-site workflows.
Why it works: Simple language lowers first-contact friction. Operator language gives the reader somewhere real to climb.
Failure mode: If too much advanced machinery appears at once, the mechanism disappears behind vocabulary.
Repair move: Use a ladder: word, example, failure, workflow, protected boundary, verification.
Practice move: Rewrite one advanced AI concept as a hallway explanation, then as an operator checklist.

### Use blunt language to force artifact clarity
Vague product language gets cut quickly when it does not point to a real action, route, source, or receipt.
Looks like: Phrases that sound strategic but do not help the user are removed or replaced with concrete content and working buttons.
Why it works: AI products are already full of vague confidence. Direct language pushes the work back to evidence.
Failure mode: The urgency can sound like impatience when the missing artifact has not been named yet.
Repair move: Translate frustration into a defect sentence: this label implies X, but the product only does Y.
Practice move: Take a vague AI dashboard sentence and rewrite it as a testable product promise.

## Operator Patterns: What To Watch
### Scope pressure can outrun priority
More content, more polish, more integrations, and more tests are often right. But without ordering, everything becomes urgent.
Looks like: Requests stack dramatic content, live research, cross-site integrations, e2e tests, focus mode, and polish into one moving target.
Why it works: The ambition catches shallow work. It prevents a small green slice from masquerading as product quality.
Failure mode: The next best action gets blurry, and useful work risks turning into broad motion.
Repair move: Force a three-lane board: must ship now, should expose publicly, and should queue as next content.
Practice move: Sort ten improvement ideas into now, next, and later based on user impact and evidence risk.

### Deploy cadence can get ahead of local proof
Fast deploys are useful, but the team must still separate local source, preview, production, and live alias evidence.
Looks like: A deployed update may become the new baseline while the next request assumes every current change is already live.
Why it works: Fast iteration keeps momentum and gets real surfaces in front of users.
Failure mode: People can talk about live behavior, local behavior, and intended behavior as if they are the same thing.
Repair move: State the evidence level in every report: local fixed, tests green, build green, preview deployed, production verified.
Practice move: Label five claims by evidence level before deciding whether a feature is truly done.

### Content volume can crowd the signal
A lot of good content can still become hard to scan if the page does not guide attention.
Looks like: Research trends, tips, demos, source ledgers, platform tips, entries, and integrations all compete for first-read attention.
Why it works: Depth is the product's advantage; shallow content would underserve serious readers.
Failure mode: The reader treats everything as equally important and misses the one AHA moment they needed.
Repair move: Make every dense section answer: start here, inspect this, try this, stop here.
Practice move: Take a dense section and mark the single first action a new reader should take.

### The platform standard can flatten product identity
A strong shared contract is good, but not every site should feel like the same observability dashboard.
Looks like: Route compliance, MCP, trust policy, trace context, and discovery can become louder than the actual product job.
Why it works: Standard surfaces make the network agent-operable and testable.
Failure mode: An insight notebook, game, docs site, and webhook lab can start sounding like the same control panel.
Repair move: Put the product purpose at the top of every pass, then wrap the standard around that purpose.
Practice move: Compare two sites and write one sentence that must stay different before listing the shared routes.

### Implicit context can become a hidden burden
When the operator says continue, they often know the whole context. The agent may need to reconstruct it from repo state, memory, live behavior, and prior instructions.
Looks like: A short follow-up can carry a lot of implied scope: latest deployment state, product tone, platform contract, and unresolved gaps.
Why it works: Short commands preserve momentum and avoid restating what should already be known.
Failure mode: If the agent reconstructs the wrong context, it can optimize an older goal or miss the current priority.
Repair move: Add one line when redirecting: current target, unacceptable failure, and what must be true when done.
Practice move: Turn a vague continue request into a one-sentence acceptance target without losing urgency.

### More can outrun source quality
More research, tips, and insights are valuable only when each claim has a job and a source boundary.
Looks like: The request for more content can tempt the system to summarize too broadly or treat trend noise as durable truth.
Why it works: A broad radar catches emerging patterns before they become obvious.
Failure mode: Quantity can make weak sources look strong because the page feels substantial.
Repair move: Use a source ledger: source, what it proves, what it cannot prove, and what should be rechecked later.
Practice move: Take three sources and state what each one can prove before writing any trend summary.

## Operator Repair Loop
1. Name the job
Move: Start with the site's actual product purpose before adding platform surfaces.
Evidence: A one-sentence role that would still make sense to a first-time user.
Watchout: Skipping straight to routes, dashboards, or MCP because those are familiar.

2. Expose the contract
Move: List public reads, protected actions, discovery files, and no-fake-data boundaries.
Evidence: OpenAPI, llms, trust policy, route inventory, and UI labels agree.
Watchout: A button or card implies an action that the code does not perform.

3. Write the dense field note
Move: Add real explanations, AHA moments, watchouts, and exercises that ramp from first contact to operator judgment.
Evidence: A reader can say what changed in how they inspect the system.
Watchout: More words without a clearer next action.

4. Prove the surface
Move: Run typecheck, lint, unit tests, build, and browser checks across desktop and mobile.
Evidence: Commands run and failures either fixed or stated plainly.
Watchout: Treating build green as proof that the product is usable.

5. Mark the remaining gap
Move: Separate local fixed, preview deployed, production verified, and next remediation.
Evidence: The final report says what changed and what remains.
Watchout: Letting momentum blur what is actually live.

## Operator Prompts
### Copy the good operator move
Use when: You want a reader or agent to imitate the strongest parts of this working style.
Prompt:
- Study this workflow and extract the operator moves worth copying.
- For each move, return: name, visible behavior, why it works, evidence required, and one practice move.
- Do not praise vibes. Only include moves that change product quality or verification quality.
Output contract:
- strengths: array of named patterns
- evidence: artifact or behavior supporting each pattern
- practice: one transfer exercise per pattern

### Name the drag without softening it
Use when: You need a candid review of where an ambitious operator style creates risk.
Prompt:
- Study this workflow and name the drag patterns.
- For each drag pattern, return: what triggers it, why it is understandable, how it breaks work, and the smallest repair move.
- Be direct, but do not psychologize. Keep it about observable workflow behavior.
Output contract:
- dragPatterns: array of observable risks
- repairMove: concrete action the operator or agent can take next
- boundary: what not to infer from the evidence

### Turn continue into proof
Use when: The instruction is short, urgent, and packed with implied context.
Prompt:
- The user said continue. Reconstruct the current goal from repo state, prior instructions, and live surface evidence.
- Return the next three proof-producing actions, the files/routes affected, and the command chain that will prove completion.
- If assumptions are necessary, state them before acting.
Output contract:
- goal: current target in one sentence
- proofActions: exactly three actions
- verification: commands or live checks
- risk: what could be stale or ambiguous

## Entries
# Use a mind microscope, not mind reading
Kind: Method
Tag: interpretability
Published: 2026-06-30
The phrase seeing inside AI is useful only if we keep it honest. We are not watching a human-like inner monologue. We are reading instruments that expose partial evidence about how an answer formed.
Token probabilities show fragile words. Attention can show routing clues. Probes can show where a concept is readable. Sparse features can separate tangled activation directions. Activation patching and circuit tracing can test whether a component actually changes behavior.
The AHA for readers is that no single view is the mind. The strongest review layers several instruments and then asks what changed in the output when the internal signal changed.
## Aha Moment
Seeing inside AI means layering instruments, not pretending one transcript reveals the whole hidden computation.

## Try This
Take one model answer and label which parts you could inspect with logprobs, attention, feature probes, circuit tests, and external receipts.

## Watch For
- Treating chain-of-thought as the literal hidden cause
- One visualization presented as the whole explanation
- Internal signals trusted more than external proof
Note: A signal is a clue. A causal intervention is stronger evidence.

# Circuit tracing gives hypotheses before certainty
Kind: Method
Tag: interpretability
Published: 2026-06-30
Modern interpretability is moving from screenshots of attention to causal stories about features and pathways. That is progress, but it also creates a new trap: the story can look more complete than the evidence.
A good field demo should label each layer: observed output, suspected feature, traced path, causal test, and external receipt. Readers should be able to say which part is a clue and which part actually changed the model's behavior.
This habit prevents interpretability theater. The beautiful graph earns trust only when it survives a behavioral check.
## Aha Moment
A circuit trace becomes persuasive when changing the suspected component changes the behavior you care about.

## Try This
Write one circuit-tracing claim as a hypothesis, then name the intervention or output check that would test it.

## Watch For
- Beautiful attribution paths with no behavioral intervention
- Feature labels that sound more precise than the evidence
- Ignoring alternative circuits that could explain the same output

# Frame agents as loops with checks
Kind: Method
Tag: agents
Published: 2026-06-30
People often hear agent and imagine a little worker with judgment. A clearer frame is a controller. It reads state, chooses one next step, takes an allowed action, checks the result, then decides whether to continue or stop.
The loop must have boundaries: what the agent may read, which tools it may call, when it must ask for PLATPHORM_API_KEY, and which check ends the run. Without those boundaries, a loop can amplify the first wrong assumption.
The AHA is that autonomy is not the absence of control. Good autonomy is visible control repeated safely.
## Aha Moment
The check step is what separates an agent loop from a repeated confident guess.

## Try This
Rewrite one one-shot AI answer as observe, decide, act, check, reflect, and require a visible artifact for every step.

## Watch For
- Loops with no stop condition
- Reflection that never changes the next action
- Retry logic that repeats the same failing move

# OpenClaw is a toolchain you can inspect
Kind: Tool
Tag: agents
Published: 2026-06-30
For this notebook, OpenClaw should be presented as an OpenClaw-compatible planning and validation lens for PlatPhorm services. That means public-safe discovery and dry-run explanation are in scope. Mutation, remediation, replay, and durable writes stay behind PLATPHORM_API_KEY on the owning service.
A reader should be able to point at every stage: capability, schema, target id, expected effect, protected boundary, verifier, and receipt. If any stage is missing, the chain is not ready to execute.
The AHA is that agent readiness is earned one artifact at a time.
## Aha Moment
OpenClaw readiness is a chain of evidence: capability, schema, dry run, boundary, verifier, and receipt.

## Try This
Draft an OpenClaw-style dry run for one remediation action and mark every public read versus protected action.

## Watch For
- Planning language that implies execution already happened
- Missing target ids or schema validation
- Protected actions shown without PLATPHORM_API_KEY boundary
Note: The learning value is in the visible chain, not in pretending a protected action already ran.

# A webhook is an event plus proof
Kind: Method
Tag: tooling
Published: 2026-06-30
The plain-language explanation is simple: one system rings another system's doorbell. The advanced version is where safety lives: who rang, what event type, what fields, what timestamp, what signature, whether it is a duplicate, and whether replay is allowed.
A public framing demo can show contract, sign, deliver, verify, retry, and replay states without sending real events or exposing secrets. Raw secrets, live event creation, replay, and durable endpoint changes belong behind the protected service boundary.
The AHA is that webhooks are receipts for cause and delivery. If the receipt fails, business logic should not run.
## Aha Moment
A webhook should be trusted only after the event contract, signature, idempotency, and delivery proof line up.

## Try This
Design a sample webhook event and list the exact reasons a receiver should reject it before running business logic.

## Watch For
- Raw secrets copied into public demos
- Replay treated as harmless because it is a retry
- Event summaries accepted without signature or schema proof

# Write prompts as portable contracts
Kind: Tool
Tag: prompting
Published: 2026-06-30
OpenAI, Claude, Gemini, and strong open-weight models differ in features and behavior. The portable layer is the part every capable model can understand: what job it is doing, which sources count, what output shape is required, what is out of bounds, and how to say unknown.
A cross-model prompt should be boring enough to test. If two models receive the same contract, you can compare schema following, receipt quality, uncertainty handling, and overreach. If you rewrite the prompt for each model, you lose the comparison.
The AHA is that model routing starts with comparable failures, not brand preference.
## Aha Moment
A portable prompt is the contract all models can be graded against, not the phrase that sounds best for one model.

## Try This
Write one prompt with role, task, source ledger, constraints, output schema, uncertainty, and stop rules, then compare two model outputs field by field.

## Watch For
- Model-specific rewrites that ruin comparison
- Schemas without unknown or blocked states
- Prompts that ask for confidence without evidence reasons

# The prompt is the whole context stack
Kind: Method
Tag: prompting
Published: 2026-06-30
Context engineering is the craft of arranging the working set so the model can use the right facts at the right time. It includes what you include, what you exclude, where you place critical constraints, and how you label sources.
More context can make answers worse if important facts are buried in the middle or mixed with low-value filler. Readers should practice source maps before summaries so the evidence has structure before the prose begins.
The AHA is that ignored context is often a design failure, not a mysterious model mood.
## Aha Moment
The model may ignore a fact because the context design buried, diluted, or mislabeled it.

## Try This
Build a source map before one summary and move the highest-risk constraint to the opening and closing of the context.

## Watch For
- More context used as a substitute for better context
- Critical facts placed in the soft middle
- Retrieved chunks pasted without ranking what they prove

# Benchmarks are maps of failure, not trophies
Kind: Method
Tag: evals
Published: 2026-06-30
A benchmark score can be useful, but the examples behind the score are usually more useful. They show what kind of task the model mishandled and what local fixture you should add before shipping.
Long-task evals are especially important for agents because short answers hide drift. A model may solve one step and still lose the goal after thirty minutes of tool use, retries, and partial failures.
The AHA is that a public benchmark should become a local test, not a marketing badge.
## Aha Moment
The most valuable benchmark output is often the broken example you can turn into a local fixture.

## Try This
Take one public benchmark-style failure and convert it into a small product-specific eval with expected safe behavior.

## Watch For
- General scores used as production guarantees
- Short-answer tests used for long-running agents
- Benchmark examples ignored after the headline score

# Reasoning is search plus checking
Kind: Method
Tag: reliability
Published: 2026-06-30
The practical mental model is search. The model tries possible routes, keeps or discards candidates, and returns the answer that survives its internal process. That does not guarantee truth; it only changes how the answer was produced.
Use reasoning models when uncertainty is high, task horizon is long, or the cost of a wrong first answer is meaningful. Then judge the result by tests, sources, calculations, or tool receipts instead of by the elegance of the reasoning.
The AHA is that deliberation buys opportunity for correction, not immunity from error.
## Aha Moment
Extra reasoning time creates chances to search and correct, but receipts still decide whether the answer is usable.

## Try This
Compare a fast answer and a slower reasoned answer, then score both only by tests, sources, calculations, or tool receipts.

## Watch For
- Verbose reasoning rewarded when evidence is weak
- Latency spent on low-risk tasks
- Reasoning traces treated as proof instead of process

# Prefer tool protocols over one-off adapters
Kind: Method
Tag: agents
Published: 2026-06-30
One-off adapters are easy to start and hard to audit. Protocol surfaces such as MCP and agent interoperability contracts give the model and the operator a shared language for what exists, what inputs are valid, and what authority is required.
The important framing move is to separate discovery from execution. A model may inspect a tool publicly without being allowed to run a protected action.
The AHA is that a protocol is not bureaucracy. It is how you make tool authority visible before the model uses it.
## Aha Moment
Protocol discovery makes tool authority visible before an agent asks to use it.

## Try This
For one tool, write the discovery metadata, input schema, public read actions, protected actions, and error states before any call.

## Watch For
- Generic tools that do not serve the site's purpose
- Execution allowed before schema inspection
- Custom adapters with undocumented auth behavior

# Prompt injection is a boundary problem
Kind: Method
Tag: security
Published: 2026-06-30
A retrieved page, webhook body, document, or user note may contain text that sounds like instructions. The model needs a rule that this text is content to analyze, not authority to obey.
The safer architecture binds tools to schemas, trusted domains, explicit user intent, protected-action checks, and refusal states. The prompt can help, but the product boundary has to carry the load.
The AHA is that security lives where text becomes action.
## Aha Moment
Prompt injection becomes dangerous when untrusted text crosses from content into command authority.

## Try This
Place a harmless fake instruction inside retrieved content and make the model classify it as content, not a command.

## Watch For
- Untrusted pages allowed to override system or user intent
- Tools exposed without least-authority checks
- Warnings in prompts used instead of product boundaries

# Build a source ledger before the answer
Kind: Tool
Tag: verification
Published: 2026-06-30
A ledger is smaller than a report and more useful than a pile of links. It turns sources into jobs. One source may prove a date, another may prove a route contract, and another may only give background.
For research trends, the ledger prevents overclaiming. A lab blog can support what the lab announced. A paper can support its tested claim. A local platform route can support current contract shape only after it is inspected.
The AHA is that evidence becomes stronger when each source has a job description.
## Aha Moment
A source ledger gives every source a job before the model starts blending evidence into prose.

## Try This
Create a four-column ledger: source, what it proves, what it cannot prove, checked-at date, then write the answer from the ledger.

## Watch For
- Links collected without claim-level purpose
- Old sources used for facts likely to drift
- Background sources treated as direct proof

# Explain it like plain language, then tighten it
Kind: Method
Tag: field
Published: 2026-06-30
A strong AI walkthrough should pass the hallway test: a reader can explain the idea to someone else without copying a paragraph. Once that works, tighten the language into the operating version.
For example: plain version, the model needs receipts for important claims. Operator version, every material claim needs a source pointer, calculation, trace id, command output, or explicit assumption before it can drive a decision.
The AHA is that plain language and expert language are not enemies. Plain language gives the reader a handle. Expert language gives the handle precision.
## Aha Moment
Simple language is the ramp, not the ceiling. The expert version lands faster after the reader can repeat the plain one.

## Try This
Rewrite one advanced AI idea twice: first as a plain-language explanation, then as an operator checklist with evidence and boundary.

## Watch For
- Simple explanations that remove the decision point
- Expert wording introduced before the reader has a handle
- Definitions that sound smart but do not change the next action
Note: A clear first explanation makes the advanced one land faster.

# Inspect one claim before judging the whole answer
Kind: Method
Tag: verification
Published: 2026-06-30
A model answer can be partly right, partly unsupported, and partly misleading at the same time. Whole-answer judgment hides that mixed state.
Choose the claim with the highest impact or the weakest evidence. Then ask: what is the source, what exactly does it say, does it support this sentence, and what would change my confidence?
This is the first verification habit most readers need. It turns AI review from opinion into evidence work.
## Aha Moment
One inspected claim can reveal whether the whole answer is evidence-shaped or just confidence-shaped.

## Try This
Pick the highest-impact sentence in a model answer and write the exact receipt that would prove or disprove it.

## Watch For
- Judging the entire answer by tone
- Checking easy claims while skipping the risky one
- Accepting a source that only supports nearby background

# Ask artifact questions, not vibe questions
Kind: Tool
Tag: field
Published: 2026-06-30
A vibe question sounds like: is this good? An artifact question sounds like: which source supports this claim? Which field is missing? Which selector is too broad? Which trace header should follow the request?
Readers ramp faster when every question has an object they can inspect. It lets them disagree productively because they are arguing over evidence, not taste.
For advanced work, artifact questions become test cases. If the group cannot point at the evidence, a release gate or agent workflow should not pretend it can.
## Aha Moment
A review gets sharper when the question points to an artifact everyone can inspect together.

## Try This
Replace one vague question like is this good with three artifact questions about a source, field, route, or tool argument.

## Watch For
- Questions that invite preferences instead of evidence
- Artifacts hidden behind summary language
- Readers unable to point at the same object

# Sort public, protected, and private before touching tools
Kind: Method
Tag: platform
Published: 2026-06-30
Public read means safe inspection: docs, feeds, sitemaps, health summaries, public entries, and read-only metadata. Protected action means mutation, ingestion, report generation, replay, or workflow execution that requires PLATPHORM_API_KEY. Private data means secrets, cookies, credentials, private traces, raw personal data, or anything not meant for public output.
The sort should happen before the model sees the task. If the action is protected, the UI should say so. If the data is private, it should not be copied into public prompts, screenshots, feeds, or discovery files.
This is a simple field exercise and a serious operator habit. It prevents a helpful assistant from becoming a careless automation system.
## Aha Moment
Most automation mistakes become visible if you sort public read, protected action, and private data before the tool runs.

## Try This
Take five routes or UI actions and label each one public read, protected action, or private data before designing the workflow.

## Watch For
- Public dashboards that imply public write access
- Protected actions presented without an auth boundary
- Private data copied into public prompts or discovery files

# Write the safe agent plan before the agent acts
Kind: Method
Tag: agents
Published: 2026-06-30
The agent plan should be boring. Boring is good. It names allowed read surfaces, protected actions, stop conditions, target ids, expected changes, and verification after the run.
A plan without boundaries is not autonomy; it is ambiguity. A plan without proof is not execution; it is a story about execution.
Readers can practice this safely with dry runs. They design the agent's allowed moves, then inspect whether the proposed tool call stays inside the plan.
## Aha Moment
An agent is easier to trust when its allowed moves, stop rules, and proof requirements are visible before execution.

## Try This
Write a dry-run agent plan with allowed reads, forbidden actions, stop conditions, target ids, and verification evidence.

## Watch For
- Agent plans that skip target ids
- Stop conditions described only after a failure
- Execution summaries with no artifact proving what happened

# A definition is not proof
Kind: Tip
Tag: dictionary
Published: 2026-06-30
This confusion appears constantly in AI learning. A reader finds a clean definition for retrieval relevance and then treats that definition as evidence that a retrieved answer is relevant.
The definition answers: what does this term mean? The receipt answers: what supports this claim here? Both matter, but they do different jobs.
The advanced move is to pair them: define the term, inspect the artifact, then ask which receipt proves the artifact meets the definition.
## Aha Moment
Definitions stabilize language. Receipts prove claims. Mixing those jobs creates false confidence.

## Try This
For one AI term, write the definition question and the proof question separately, then answer both with different sources.

## Watch For
- A clean definition treated as evidence for a live claim
- Receipts judged without agreeing on the term
- Dictionary links used where a current source is required

# Every button needs a contract
Kind: Tool
Tag: integration
Published: 2026-06-30
Buttons are tiny contracts. Bad buttons imply writes that never happened, automation that is not wired, or authority the user does not have.
A good button tells the truth in the label, sends the user to the right surface, and leaves a source trail. A better button also exposes the public JSON or receipt behind the content.
This is why a learning notebook should make Try this actions real: copy the exercise, open the learning surface, and expose the source entry.
## Aha Moment
A button is a product promise. If the action is only copy, link, or read JSON, the label should say exactly that.

## Try This
Audit three UI buttons and write what each one actually does, what it does not do, and what proof the user receives.

## Watch For
- Buttons that imply a write happened
- Protected actions without PLATPHORM_API_KEY
- No source link behind generated or instructional content

# Build the source map before the summary
Kind: Method
Tag: verification
Published: 2026-06-30
A source map is a small table: source, what it owns, what it proves, what it cannot prove, and when it was checked. For PlatPhorm surfaces, that might include network graph, sitemap index, robots, llms files, OpenAPI, MCP manifest, and RSS.
Without a source map, summaries become confident blends of current facts, stale memory, and assumptions. With a source map, the model has to show where each claim came from.
The AHA is that summarization becomes safer when the evidence list exists before the prose.
## Aha Moment
A summary becomes safer when the source map exists before the prose starts.

## Try This
Create a source map with source, what it proves, what it cannot prove, and checked date before asking AI for a summary.

## Watch For
- Summaries that blend live evidence with stale memory
- Sources used outside what they can prove
- No checked-at signal for drift-prone facts

# Turn the mistake into the AHA
Kind: Tip
Tag: field
Published: 2026-06-30
If every demo is polished, readers learn to admire the guide. If a safe mistake is visible, readers learn how to recover.
Use mistakes that reveal a transferable pattern: a citation that is adjacent but not supportive, a tool selector that is too broad, a definition that is useful but not proof, or a status card with no source.
The AHA should be the repaired habit: inspect one claim, demand a receipt, sort the boundary, or ask an artifact question.
## Aha Moment
A safe mistake can frame the repair habit more clearly than a flawless demo.

## Try This
Design one harmless failure case, let readers diagnose it, then write the habit that would catch it next time.

## Watch For
- Mistakes that embarrass readers instead of exposing a pattern
- Repairs performed silently by the guide
- AHA moments that do not name a future action

# Use red, yellow, green for action readiness
Kind: Tool
Tag: field
Published: 2026-06-30
The scale works because it ties confidence to action. Green does not mean perfect. It means the evidence is good enough for the next low-risk step. Yellow means uncertainty changes what happens next. Red means the system should stop or ask for approval.
Use it on model answers, tool previews, citations, glossary definitions, platform routes, and agent plans. The reader's job is to name the color and the reason.
As readers get stronger, replace color with richer states: unknown, partial, protected-required, needs-review, blocked, or ready.
## Aha Moment
Readiness labels matter when they change what happens next.

## Try This
Score one model answer green, yellow, or red and require the reason plus the next action for that color.

## Watch For
- Green used to mean perfect instead of ready for the next step
- Yellow states with no inspection plan
- Red states that stop without explaining the boundary

# Frame the term before the trick
Kind: Method
Tag: field
Published: 2026-06-30
The fastest field failure is showing the advanced move while readers are still guessing what the nouns mean. A prompt fixture, dry run, receipt, confidence bucket, and abstention branch each need a plain-language definition before the walkthrough begins.
Use the Dictionary connection as a framing move: introduce the term, point to the public definition surface, ask readers to predict the failure mode, and only then run the demonstration.
This changes the guide's job from explaining a magic trick to naming a repeatable mechanism. Once the term is stable, readers can transfer it to a new model, tool, or workflow.
## Aha Moment
The vocabulary is part of the interface. If the term is unstable, the walkthrough is unstable.

## Try This
Start each framing block by naming three terms readers must understand before they watch the model behavior.

## Watch For
- Readers copying a workflow while using the key term differently
- Demos that rely on undefined platform language
- Definitions introduced only after confusion appears
Note: A shared glossary lowers the cost of every later explanation.

# Use Dictionary as the concept receipt
Kind: Method
Tag: dictionary
Published: 2026-06-30
Concept receipts are different from fact receipts. A fact receipt supports a claim about the world. A concept receipt stabilizes the language used to discuss the claim.
This matters in AI framing because terms like context window, retrieval, grounding, confidence, eval, and tool call get used casually. Casual vocabulary creates accidental disagreement.
A good integration links to Dictionary for the term, links to the notebook for the operating habit, and links to Notes for the reader's private or public reflection workflow.
## Aha Moment
A Dictionary link can be a concept receipt that proves the group is sharing a meaning before debating evidence.

## Try This
Attach a Dictionary definition to each specialized term in a field slide, notebook entry, or review checklist.

## Watch For
- Treating a term definition as proof of a factual claim
- Definitions whose domain does not match the walkthrough
- Glossary links that point to search pages instead of concrete terms

# Turn confused notes into dictionary candidates
Kind: Method
Tag: notes
Published: 2026-06-29
A field generates terminology demand in real time. If readers keep asking what a phrase means, that phrase belongs in a glossary queue.
Notes is the capture layer: readers write the confusion in their own language. Dictionary is the stabilization layer: the group turns that recurring confusion into a reviewed definition.
The important boundary is human approval. The model can suggest a term, summary, or definition draft, but the published Dictionary meaning should be reviewed before it becomes shared framing infrastructure.
## Aha Moment
Repeated reader confusion is not noise. It is demand for a clearer shared definition.

## Try This
Collect fuzzy reader phrases in Notes, group repeated phrases, and promote only reviewed candidates to Dictionary.

## Watch For
- Publishing model-suggested definitions without human review
- Losing repeated reader confusion after the group ends
- One-off answers that never improve the shared glossary

# Run the explain, inspect, repair loop
Kind: Method
Tag: field
Published: 2026-06-29
Explaining a concept gives readers vocabulary. Inspecting an artifact gives them evidence. Repairing a failure gives them agency.
For prompt reliability, the artifact might be a bad answer. For tool safety, it might be a dry-run plan. For verification, it might be a citation that points near the claim but not under it.
The loop should end with an AHA checkpoint: what changed in how the reader sees the model, not just what command they copied.
## Aha Moment
Readers remember a concept when they see the mechanism, inspect evidence, and repair a failure themselves.

## Try This
Frame one topic with three beats: explain the mechanism, inspect a concrete artifact, then repair the broken part.

## Watch For
- Slides that explain without giving readers an artifact to inspect
- Live demos with no repair step
- Repairs that happen silently in the guide's head

# Climb from word to workflow
Kind: Tip
Tag: field
Published: 2026-06-28
The word rung names the concept. The example rung shows it in a clean case. The failure rung shows why it matters. The workflow rung turns it into a repeatable decision.
Dictionary helps with the word rung. The notebook helps with the example and failure rungs. Notes helps readers turn the workflow rung into their own operating habit.
If a walkthrough stops at a definition, it is reference material. If it reaches workflow, it becomes practice.
## Aha Moment
A term becomes operational only when the reader can use it to choose the next move.

## Try This
For one concept, write the word, a clean example, a failure example, and the workflow decision it changes.

## Watch For
- Definitions that never reach a practical example
- Examples that do not show the failure mode
- Readers who can repeat the word but cannot act on it

# Put the glossary before the demo
Kind: Tip
Tag: dictionary
Published: 2026-06-28
Before a retrieval demo, define retrieval, relevance, chunk, contradiction, and receipt. Before a tool demo, define dry run, target id, protected action, and rollback.
The definitions do not need to be long. They need to be visible and shared.
This is where a Dictionary link does real work. It gives the group a stable reference outside the slide, the notebook entry, and the guide's memory.
## Aha Moment
The demo is where readers should watch behavior, not decode vocabulary under pressure.

## Try This
List the walkthrough glossary before the first live output and link each term to Dictionary when a public definition exists.

## Watch For
- Cognitive overload from new terms during the demo
- Guide-specific jargon that has no shared reference
- Terms explained differently in slides, notes, and discussion

# Keep glossary publishing human-approved
Kind: Method
Tag: governance
Published: 2026-06-27
The risk is subtle. A bad definition does not only misstate one answer; it trains future readers to reason with the wrong frame.
A human approval queue catches invented authority, missing caveats, audience mismatch, and definitions that are technically correct but useless in practice.
The pipeline is still fast. Notes can stage the reader language, AI can draft the candidate, Dictionary can expose the reviewed term, and the notebook can link the term back into walkthroughs.
## Aha Moment
A bad shared definition compounds because future walkthroughs and workflows inherit it.

## Try This
Stage candidate glossary definitions as drafts, require human approval, and include source or scope notes before publishing.

## Watch For
- Model-written definitions treated as approved knowledge
- No provenance for why a definition changed
- Technically correct definitions that miss the field audience

# Use integration cards instead of vague platform links
Kind: Tool
Tag: integration
Published: 2026-06-27
Dictionary owns public definitions and concept lookup. Notes owns learning capture, term detection, and note analysis previews. This notebook owns walkthroughs, entries, and framing flows.
A good integration card states the handoff in plain language: open the definition, capture a note, preview term detection, or review a workflow template.
It should also name the boundary. If this site is not posting to Dictionary or Notes, the UI should not imply that a write happened.
## Aha Moment
A cross-site link is safer when it says what the target owns and what this page will not do.

## Try This
Write each integration card with owner, public reads, protected actions, and handoff links before adding buttons.

## Watch For
- Buttons implying a write happened when only navigation occurred
- Protected actions presented as public workflows
- Generic platform links without a product-specific handoff
Note: A link without a contract is just navigation.

# Stop for AHA checkpoints
Kind: Method
Tag: field
Published: 2026-06-26
AHA checkpoints are small but powerful. They ask readers to name the shift: from answer to evidence, from prompt to contract, from tool to action, from confidence to decision.
The checkpoint should be concrete enough to grade informally. If readers cannot say what they would inspect next, the AHA moment did not land yet.
This is why the notebook includes AHA text beside entries instead of hiding it in guide notes.
## Aha Moment
The AHA moment is the measurable change in what the reader knows how to inspect next.

## Try This
Pause after each major concept and ask readers to name the first artifact they would inspect now.

## Watch For
- Readers agreeing verbally without changing their review behavior
- Checkpoints that ask for feelings instead of observable next moves
- Too many concepts between reflection points

# Use Notes as an observation log
Kind: Tip
Tag: notes
Published: 2026-06-26
Observation-first notes reduce hindsight storytelling. The reader writes the low-confidence token, unsupported claim, weird citation, or broad tool selector before explaining it away.
Those observations become better prompts for review, better term candidates for Dictionary, and better discussion material for a field.
The simple format is: I saw, I expected, I think it means, I need a definition for.
## Aha Moment
Observation-first notes preserve the evidence before the reader writes a story about it.

## Try This
Use the format I saw, I expected, I think it means, and I need a definition for after every demo.

## Watch For
- Conclusions recorded without the artifact that caused them
- Surprising behavior erased by a polished summary
- Term confusion captured nowhere reusable

# Check provenance before framing a term
Kind: Method
Tag: dictionary
Published: 2026-06-25
Definitions have context. A release-engineering meaning of gate is not the same as a field meaning or a safety-policy meaning.
Before framing from a term, check whether the definition names the domain, source, and intended audience. If the provenance is weak, treat it as a draft or discussion prompt.
This keeps Dictionary integration honest: it supports shared language, but it does not replace judgment.
## Aha Moment
A definition is scoped evidence. It helps only when its source and domain match the walkthrough.

## Try This
Before framing a term, inspect whether the definition names the domain, source, and intended audience.

## Watch For
- One domain's meaning reused in a different workflow
- Definitions with no source or scope signal
- Treating Dictionary as a substitute for current behavior checks

# Use Socratic eval prompts in group
Kind: Tool
Tag: evals
Published: 2026-06-25
Socratic eval prompts frame readers to inspect systems instead of merely reacting to outputs. The question shapes the habit.
Good field eval prompts are short, reusable, and attached to visible artifacts. They work on a model answer, a tool plan, a note, a definition, or a citation list.
The strongest prompt is often: what would you need to see before you let this action run?
## Aha Moment
The question shapes the habit. Ask about failure, missing receipts, and next evidence instead of asking if it is good.

## Try This
Give readers four prompts: what fails first, what receipt is missing, what term is undefined, and what evidence changes this.

## Watch For
- Eval questions that invite taste instead of inspection
- Readers scoring outputs without citing artifacts
- Questions that do not transfer to another workflow

# Map misconceptions as content gaps
Kind: Method
Tag: field
Published: 2026-06-24
A misconception map turns framing friction into backlog. If readers confuse confidence with correctness, add an entry. If they confuse retrieval with relevance, add a definition. If they over-trust citations, add a failure case.
The map should include the mistaken belief, the correction, the artifact that reveals it, and the next content improvement.
This is how a field notebook becomes a living curriculum instead of a static list of tips.
## Aha Moment
Misconceptions are content backlog with evidence attached.

## Try This
Track repeated misunderstandings as missing entries, missing definitions, missing examples, or missing exercises.

## Watch For
- Answering the same confusion repeatedly without improving content
- Misconceptions blamed on readers instead of unclear material
- No owner for converting group friction into curriculum updates

# Use full-screen mode for shared attention
Kind: Tool
Tag: field
Published: 2026-06-24
Full-screen mode is not a marketing hero. It is a field tool: large type, clear current walkthrough, one AHA checkpoint, and one action at a time.
The best version preserves the operating surface. Readers can still open the related entry, check the Dictionary term, or capture a note after the slide lands.
Use it when shared attention matters: kickoff, recap, live demo, or review.
## Aha Moment
Full-screen mode is useful when it narrows the room to one concept, one artifact, and one next action.

## Try This
Use full-screen mode for kickoff, live demo, recap, or AHA checkpoint, then return to the notebook for details.

## Watch For
- Presentation mode that hides links readers need after group
- Large decorative screens with no workflow action
- Too much text on a shared framing slide

# Watch for glossary drift
Kind: Method
Tag: governance
Published: 2026-06-23
AI vocabulary moves quickly. A definition that was clear during one model generation can become misleading after tools, providers, or platform boundaries change.
Drift review asks whether the term still matches current behavior, whether the examples still reproduce, and whether the protected/public boundary changed.
The repair is small but important: update the definition, update the entry, and update any field checkpoint that depends on it.
## Aha Moment
Shared vocabulary is a dependency, and dependencies drift.

## Try This
Review field terms after model, tool, or platform boundary changes and update Dictionary plus notebook links together.

## Watch For
- Definitions that describe a previous platform state
- Entries framing examples that no longer reproduce
- Protected/public boundaries changing without glossary updates

# Turn entries into self-quiz cards
Kind: Tool
Tag: field
Published: 2026-06-23
Recall asks for the idea. Artifact asks readers to use the idea on a concrete output. Transfer asks whether they can spot the same pattern in a different workflow.
For receipts: what is a receipt, which claim lacks one, and where else would you require one? For dry runs: what is the action preview, which selector is too broad, and what protected action should stop?
Self-quiz cards make the notebook useful between groupes, not only during live framing.
## Aha Moment
Learning sticks when recall, artifact inspection, and transfer all work.

## Try This
Write one recall question, one artifact question, and one transfer question for each high-value entry.

## Watch For
- Quizzes that only ask for memorized definitions
- No concrete artifact in the question
- Readers passing the quiz but failing to transfer the habit

# Score retrieved context before you trust the answer
Kind: Method
Tag: verification
Published: 2026-06-28
RAG failures often look like model failures because the model speaks with the confidence of the context it was given. The first move is to separate retrieval quality from generation quality.
Give each chunk three labels: directly answers, adjacent background, or contradiction. The answer should cite directly answering chunks first and name contradictions instead of smoothing them away.
A good retrieval review is boring and mechanical. It asks whether the evidence was even in the room before debating whether the model reasoned over it correctly.
## Aha Moment
Retrieval only proves that documents were fetched. It does not prove the fetched text answers the question.

## Try This
Label each retrieved chunk as direct support, background, contradiction, or irrelevant before reading the final answer.

## Watch For
- Answers citing adjacent background as if it were direct support
- Contradictory chunks hidden by smooth synthesis
- High source count used as a substitute for relevance
Note: Retrieved is not the same thing as relevant.

# Design the output schema before the prompt
Kind: Tip
Tag: reliability
Published: 2026-06-24
A schema makes ambiguity visible. If you cannot name the field, type, and allowed absence state, you probably do not know what you are asking the model to produce.
Start with the downstream consumer. Does it need a label, a span, a confidence bucket, a source pointer, or an abstention? Write that contract first and then prompt the model to fill it.
The aha is that schemas are not just for machines. They force human reviewers to decide what counts as complete, partial, unknown, and unsafe.
## Aha Moment
A schema turns a vague prompt into an interface contract with explicit absence states.

## Try This
Write the JSON fields, allowed values, evidence pointers, and unknown states before drafting the prompt.

## Watch For
- Required fields that force the model to invent missing data
- Enums that do not include blocked, unknown, or partial
- Fields whose consumer or decision impact is unclear

# Use the eval sandwich: before, during, after
Kind: Method
Tag: evals
Published: 2026-06-20
Before-change evals establish the baseline you are trying not to lose. During-change evals are small and sharp: the five examples most likely to break because of the edit. After-change evals make sure the fix did not damage neighboring behavior.
This is especially useful for system prompt work because every instruction edit has a shadow. You improve one behavior and accidentally move refusal, tone, verbosity, or citation discipline.
Treat the eval sandwich as a change-management ritual. It is less glamorous than a benchmark, but it is the thing that keeps production behavior from drifting by inches.
## Aha Moment
The most valuable eval is often the small one you run while the prompt is still being edited.

## Try This
Run a baseline set before the change, five targeted fixtures during the change, and a regression set after.

## Watch For
- Only testing the behavior you intended to improve
- Skipping refusal and edge fixtures after style changes
- Treating one green benchmark as proof of production safety
Note: The middle eval is where most prompt regressions get caught.

# Dry-run tool calls before letting them mutate anything
Kind: Method
Tag: tooling
Published: 2026-06-14
Tool use changes the risk profile of a model. The model can be wrong in the world, not just wrong in text. The safe pattern is preview, inspect, authorize, execute, verify.
The dry run should include exactly what will be called, what identifiers will be touched, which fields will change, and what success will look like. If the model cannot state those plainly, it should not run the tool.
The best dry runs also include a refusal branch. They say what condition would stop execution and what evidence would be needed to proceed.
## Aha Moment
A tool call should become inspectable before it becomes irreversible.

## Try This
Require action preview, target ids, expected effects, refusal conditions, and rollback notes before execution.

## Watch For
- Tool arguments that include broad or ambiguous selectors
- Mutation plans without a verification step
- The model treating permission to plan as permission to execute

# Triangulate citations instead of trusting the first source
Kind: Method
Tag: verification
Published: 2026-06-08
Citation checking should ask three questions: does the source exist, does it support the exact claim, and would another independent source phrase the same claim the same way?
When sources disagree, the disagreement is the content. Do not ask the model to hide the conflict behind a confident synthesis. Ask it to name the split and explain what would settle it.
Triangulation is not needed for every sentence. Reserve it for claims that drive decisions: dates, quotes, prices, medical or legal facts, benchmarks, and allegations.
## Aha Moment
A second independent source often reveals whether a citation supports the claim or merely resembles it.

## Try This
Triangulate dates, quotes, benchmarks, allegations, and high-impact facts before presenting them as settled.

## Watch For
- Citations that support a nearby but different claim
- Two sources that repeat the same upstream source
- Disagreements collapsed into a false consensus

# Red-team prompts with harmless mutants
Kind: Method
Tag: evals
Published: 2026-05-31
Prompt boundaries are brittle when tests are too literal. A model may refuse the obvious bad phrasing and comply with the polite version, the translated version, or the version hidden inside a roleplay.
Mutants make that brittleness visible without adding new dangerous content. They perturb the wrapper, not the underlying safety category.
Use them for both sides of the boundary: prompts that should be refused and prompts that should remain allowed. Over-refusal is a product bug too.
## Aha Moment
A boundary that only holds for one phrasing is a memorized example, not a reliable rule.

## Try This
Generate harmless variants that preserve intent while changing tone, order, indirection, language, and role framing.

## Watch For
- Safety behavior that fails when the user sounds polite
- Allowed prompts that become over-refused after hardening
- Fixtures that mutate the danger instead of just the wrapper

# Diff the meaning, not just the text
Kind: Tool
Tag: prompting
Published: 2026-05-18
Text diff says what characters changed. Semantic diff says what the model is now being asked to believe, prioritize, refuse, or produce. That is the diff that matters for behavior.
Look for modal verbs, scope words, exception clauses, examples that imply new rules, and output format changes. These are the places where small edits become big runtime differences.
A useful semantic diff ends with a test plan: which fixtures should change, which must not change, and which new example proves the edit worked.
## Aha Moment
The smallest text diff can be the biggest behavioral diff when it changes authority, scope, or obligation.

## Try This
Review prompt edits for new duties, removed constraints, softened verbs, and examples that imply new policy.

## Watch For
- May, should, and must changes hidden in wording cleanup
- Examples that contradict the written rule
- Output contracts changed without matching tests
Note: A one-word modal change can be a policy change.

# Spend latency where uncertainty is highest
Kind: Tip
Tag: reliability
Published: 2026-05-09
Reliable AI systems are often cascades. Cheap checks run first; expensive checks wake up only when the situation deserves them. This keeps the product fast without pretending every answer has equal risk.
A practical budget has triggers: low logprob margin, missing receipt, protected action, high-dollar decision, policy boundary, or user-visible publication.
The aha is that latency can be evidence-driven. You do not need to choose between always fast and always careful.
## Aha Moment
Careful does not have to mean slow everywhere. Verification can wake up only where risk wakes it up.

## Try This
Escalate to extra checks when confidence is low, evidence is weak, impact is high, or an action is protected.

## Watch For
- Uniformly expensive checks on low-risk tasks
- Fast paths that ignore missing evidence
- Latency budgets without explicit escalation triggers

# Route tasks by failure mode, not brand name
Kind: Method
Tag: reliability
Published: 2026-04-26
Model routing is usually framed as cost or quality. The more useful frame is failure. Which model omits caveats, over-refuses, invents sources, follows schemas, or handles long contexts under pressure?
Build small route tests around the jobs you actually run. A model that wins general reasoning may still be the wrong choice for structured extraction if it casually fills unknown fields.
Keep routing reversible. When model behavior shifts, the route table should be an operational setting backed by evidence, not a belief system.
## Aha Moment
Routing by failure mode is more useful than routing by reputation because every model breaks differently.

## Try This
Build a route table from observed task fixtures: extraction, synthesis, coding, refusal review, and long-context use.

## Watch For
- Leaderboards used as a substitute for local fixtures
- Model routes that cannot be changed without code edits
- Ignoring behavior drift after model updates

# Treat memory as evidence with an expiration date
Kind: Tip
Tag: verification
Published: 2026-04-18
Good memory is scoped. It records what the user decided, what a repo convention is, and which risks were already found. Bad memory stores private residue, stale status, or facts that should be rechecked.
Every memory-derived claim should answer two questions: when was this learned, and how likely is it to have changed? If the answer is 'recently unstable,' verify before acting.
The aha is that memory is not context magic. It is a cache, and caches need invalidation rules.
## Aha Moment
Agent memory is a cache. Useful caches have scope, freshness, and invalidation rules.

## Try This
Store stable decisions and preferences, but re-check drift-prone facts before acting on them.

## Watch For
- Old status presented as current proof
- Private residue stored when only workflow shape was needed
- Memory outranking live repo, source, or user evidence

# Design abstention as a first-class output
Kind: Tip
Tag: reliability
Published: 2026-04-03
Abstention fails when it feels like an error. It works when it is a designed branch with a reason, missing evidence, and next action. The user should know what changed from answerable to not answerable.
Use explicit output states: answered, partial, unknown, blocked, protected-required, and needs-review. Each state should carry what evidence exists and what evidence is missing.
This is how you prevent graceful prose from hiding a dead end. The model can stop without leaving the user stuck.
## Aha Moment
A model fills blanks when the product has no respected way to stop.

## Try This
Make unknown, partial, blocked, protected-required, and needs-review states valid outputs with next actions.

## Watch For
- Schemas where every field must be filled even when evidence is absent
- Refusals that leave the user without a useful next step
- Unknown states treated as errors instead of designed outcomes

# Use confidence buckets, not fake precision
Kind: Method
Tag: verification
Published: 2026-03-21
Precise confidence numbers look scientific and often are not. Unless they are tied to a calibrated model or measured historical error, they are mostly theater.
Buckets work better in operator workflows because they can map to action: high can publish, medium can request a receipt, low can escalate, unknown can abstain.
The bucket should cite the reason: direct source found, conflicting source found, no source found, or model-only inference. That reason matters more than the label.
## Aha Moment
Confidence is useful when it changes what happens next, not when it looks mathematically precise.

## Try This
Use high, medium, low, and unknown buckets tied to evidence type, conflict level, and review action.

## Watch For
- Percentages that are not calibrated against historical outcomes
- Confidence labels without evidence reasons
- High confidence on claims with no direct receipt

# Ask for counter-evidence before confidence
Kind: Method
Tag: reliability
Published: 2026-06-18
The useful move is not asking the model to be uncertain. The useful move is forcing it to search for the best case against itself while the answer is still provisional.
A simple pattern: draft the answer, list disconfirming evidence, name what would change the answer, then assign confidence. The confidence number becomes attached to evidence quality instead of verbal smoothness.
This is especially helpful for research summaries, legal-ish interpretation, data diagnosis, and anything where one missing constraint can flip the conclusion.
## Aha Moment
The model is usually best at self-correction before it has been rewarded for sounding certain.

## Try This
Ask for the answer, the strongest objection, what evidence would change the answer, and then confidence.

## Watch For
- Objections that restate the answer instead of challenging it
- Confidence that stays high after missing evidence appears
- A final answer that ignores its own counter-evidence
Note: Confidence after counter-evidence is more useful than confidence after a sales pitch.

# Turn every important claim into a receipt
Kind: Method
Tag: verification
Published: 2026-05-27
Models are good at moving from evidence to prose. They are also good at producing prose that looks like it came from evidence. The receipt requirement separates those two behaviors.
The receipt can be a citation, a row count, a command output line, a test name, or a declared assumption. The format matters less than the discipline: every claim has to point somewhere.
When the model cannot produce a receipt, that is not failure. That is the system doing its job and showing you where judgment is still needed.
## Aha Moment
A citation is not decoration. It is a handle you can pull to see whether the sentence is load-bearing.

## Try This
Make every claim end with one receipt: source, row, command, calculation, trace id, or explicit assumption.

## Watch For
- Receipts that point to a broad document instead of the claim
- Numbers without denominators
- Source links that support nearby context but not the sentence itself

# Keep the system prompt short enough to audit
Kind: Tip
Tag: prompting
Published: 2026-05-03
System prompts work best as durable operating constraints, not as a dumping ground for every possible preference. When the prompt grows without structure, the model starts resolving conflicts in ways you cannot predict.
Split the prompt into policy, role, output contract, and refusal boundary. Then write a small regression set that exercises each section. You will catch more problems with five targeted tests than with another page of instructions.
## Aha Moment
A system prompt is product code. If it is too long to review, it is too long to trust.

## Try This
Rewrite the system prompt into role, hard boundary, output contract, and escalation rule.

## Watch For
- Repeated style instructions that fight each other
- Hidden policy changes bundled into wording edits
- Task detail that belongs in user input rather than permanent rules
Note: If you cannot diff it cleanly, you cannot debug it cleanly.

# Use refusal boundaries as test fixtures
Kind: Method
Tag: evals
Published: 2026-04-12
Most prompt tests focus on whether the model can complete the intended workflow. That misses the other half of the product: whether it handles edge requests without over-refusing, leaking private context, or inventing authority it does not have.
Build a small fixture set with three buckets: allowed, disallowed, and allowed-with-constraints. A good model response should land in the right bucket and explain the usable next step.
This turns safety from a vibe into a regression target. When a prompt edit changes the refusal boundary, you see it immediately.
## Aha Moment
Safety quality is not whether the model refuses. It is whether it refuses at the right boundary.

## Try This
Keep allowed, disallowed, and allowed-with-constraints prompts as fixtures and run them after prompt edits.

## Watch For
- Over-refusal on harmless transformed versions
- Under-refusal when the same intent is phrased politely
- Responses that refuse but fail to offer a safe next step

# Read the logprobs, not the answer
Kind: Method
Tag: interpretability
Published: 2025-11-02
Every token a model emits comes with a probability distribution over the whole vocabulary. The decoded text only shows you the winner. The logprobs show you the race.
When you pull them back, scan for the spots where the top token barely beat the runner-up. Those are the seams. Names, dates, citations, and the first token of a refusal are where you most often find a coin-flip dressed up as a confident sentence.
A practical loop: request the top 5 logprobs per token, render them as a heatmap over the output, and read the cold spots first. That is usually where the hallucination lives.
## Aha Moment
The final text is only the winning token path. The uncertainty is still visible in the runners-up.

## Try This
Render low-margin tokens first, especially names, dates, citations, and first tokens after a policy boundary.

## Watch For
- Fluent spans built from low-confidence token choices
- Nearly tied alternatives around entities
- A confident tone hiding a fragile probability margin
Note: Confidence ≠ correctness. But low confidence is a great place to look for hallucinations.

# Make it think before it answers
Kind: Tip
Tag: prompting
Published: 2025-10-28
Autoregressive models can only condition on what they have already written. If the answer comes first, the explanation is generated to fit a token that is already locked in. It is rationalization, not reasoning.
Flip the order and the reasoning tokens become real scratch space the final answer can lean on. This is the entire mechanism behind chain-of-thought, and it is why 'answer, then explain' quietly underperforms.
## Aha Moment
Explanation after an answer often explains a commitment; reasoning before an answer can still change it.

## Try This
Put scratch work, constraints, and uncertainty checks before the final answer field.

## Watch For
- Post-hoc explanations that never revise the answer
- A final answer that appears before evidence is considered
- Reasoning sections that only paraphrase the conclusion

# BertViz for watching attention move
Kind: Tool
Tag: tooling
Published: 2025-10-19
BertViz draws the attention pattern as lines connecting tokens, split by head and layer. It runs in a notebook in a couple of lines and works on most Hugging Face transformer models.
The value is not the pretty picture, it is the habit of looking. Once you have seen a head that exclusively tracks matching brackets, you stop thinking of the model as a black box and start thinking of it as a stack of small, legible mechanisms.
## Aha Moment
Attention is not the whole mind, but it is a useful map of what information is being routed where.

## Try This
Inspect heads for repeated roles: bracket matching, quote tracking, subject agreement, or copied labels.

## Watch For
- Pretty attention maps with no behavioral test
- One dramatic head treated as the whole explanation
- Patterns that disappear after small prompt perturbations
Note: Most heads are boring. The interesting 5% are worth the dig.

# Tell it what NOT to do, sparingly
Kind: Tip
Tag: prompting
Published: 2025-10-11
Putting a word in the prompt raises its activation, even when the word is wrapped in 'do not'. The model does not have a clean negation operator the way code does; it has attention, and attention does not know the difference between mention and command.
Rewrite prohibitions as positive descriptions of the target. Instead of 'do not be formal', say 'write like you are texting a friend'. Save the hard negatives for the few rails you cannot express any other way.
## Aha Moment
Negation still puts the forbidden idea into context. Attention does not forget because grammar says not to.

## Try This
Rewrite prohibitions as the positive shape of the output you want before adding hard safety rails.

## Watch For
- Forbidden words repeated in the response
- Style instructions expressed mostly as negatives
- A model fixating on the thing you tried to suppress

# Probe for a concept with a linear classifier
Kind: Method
Tag: interpretability
Published: 2025-09-30
The setup is small: collect hidden states from a chosen layer for a labeled dataset, then fit logistic regression on top. High accuracy means the concept is linearly available at that layer.
Sweep the probe across layers and you can watch where a concept first becomes readable. Abstract properties tend to sharpen in the middle layers, which is a clue about where the model is doing the relevant work.
## Aha Moment
If a tiny linear classifier can read a concept from hidden states, the model has already organized that signal.

## Try This
Probe one concept across layers and look for where accuracy first rises above a simple baseline.

## Watch For
- High probe accuracy mistaken for causal importance
- Labels that leak through surface wording
- Layer sweeps without a held-out validation split
Note: Cheap, fast, and weirdly revealing.

# Temperature is a creativity dial, not a quality dial
Kind: Tip
Tag: prompting
Published: 2025-09-22
Temperature reshapes the probability distribution before sampling. Low temperature sharpens it toward the most likely token; high temperature flattens it so unlikely tokens get a real chance.
So the dial controls variance, not intelligence. If a task has a correct answer, variance is your enemy and you want it near zero. If you want range and surprise, turn it up and sample several times.
## Aha Moment
Temperature changes variance, not intelligence. It widens the lottery; it does not improve the tickets.

## Try This
Set temperature near zero for extraction and scoring, then raise it only when variety is the product.

## Watch For
- Creative settings used for factual extraction
- One sampled answer treated as a stable result
- Quality claims that ignore variance between runs

# Sparse autoencoders to untangle neurons
Kind: Method
Tag: interpretability
Published: 2025-09-14
The trick is over-completeness plus sparsity. You map activations into a much wider space but force only a few of those dimensions to be active at once. The pressure pushes tangled concepts apart into separate, more interpretable features.
Once you have the features you can do surgery: amplify one and watch the model fixate on a bridge, suppress another and watch a behavior fade. It is the closest thing we have to named knobs inside the network.
## Aha Moment
Sparse features give you named-ish knobs because they separate mixed neuron behavior into cleaner directions.

## Try This
Treat an SAE feature as a hypothesis, then test whether steering it changes the behavior you care about.

## Watch For
- Human-readable labels that overfit a few examples
- Feature steering without downstream behavior checks
- Polysemantic residue hiding inside a cleaner-looking feature
Note: This is roughly how the 'Golden Gate Claude' demo worked.

# Keep a tokenizer open in a tab
Kind: Tool
Tag: tooling
Published: 2025-09-03
Leading spaces, casing, and digit grouping all change how text is split, and the model only ever sees the tokens, never your characters. A trailing space can silently sabotage a few-shot format.
When something is off and you cannot explain it, paste the exact string into a tokenizer view before you theorize. The answer is in there more often than you would expect.
## Aha Moment
The model never sees your characters. It sees chunks, spaces, merges, and weird number splits.

## Try This
Paste the exact prompt, including spaces and separators, into a tokenizer before debugging prompt behavior.

## Watch For
- Leading spaces changing label tokens
- Numbers split into surprising chunks
- Few-shot examples with inconsistent separators

# Sample many, keep the majority
Kind: Method
Tag: reliability
Published: 2025-08-25
Different samples take different reasoning paths but tend to converge on the same correct answer, while the wrong answers scatter. Voting over the final answers turns that asymmetry into accuracy.
It costs you N times the compute, so reserve it for the questions that matter. A useful middle ground is to escalate to self-consistency only when the first answer looks low-confidence.
## Aha Moment
Wrong reasoning paths tend to scatter; correct reasoning paths often converge.

## Try This
Sample several answers, normalize the final claim, and inspect disagreement before majority voting.

## Watch For
- Repeated wording counted as independent evidence
- Voting over outputs that all share the same missing assumption
- Compute spent on low-value questions

# Mind the middle of long contexts
Kind: Tip
Tag: prompting
Published: 2025-08-12
Retrieval accuracy as a function of position tends to be U-shaped: strong at the start, strong at the end, sagging in the middle. A fact you need can be technically present and still effectively invisible.
Design around it. Lead with the instruction, restate the critical constraint at the very end, and keep the soft middle for material you can afford to lose.
## Aha Moment
A fact can be present and still practically invisible if it sits in the soft middle of a long prompt.

## Try This
Move critical constraints to the opening contract and repeat the single highest-risk fact at the end.

## Watch For
- Middle facts ignored while opening and closing instructions are followed
- Retrieved context pasted without ranking
- Important constraints surrounded by low-value filler
Note: Sometimes called 'lost in the middle'.

# Activation patching to find the cause
Kind: Method
Tag: interpretability
Published: 2025-07-30
Probes tell you what is represented. Patching tells you what matters. By swapping one activation and watching the effect, you move from 'this information is present' to 'this component causes the behavior'.
Sweep the patch across layers and positions to localize the circuit. The spots where patching flips the answer are the load-bearing ones; everything else is along for the ride.
## Aha Moment
Correlation says a signal is present. Patching asks whether changing that signal changes the outcome.

## Try This
Patch one layer-position activation from a clean run into a corrupted run and track whether the answer flips.

## Watch For
- Large sweeps without a clean causal hypothesis
- Output flips caused by formatting rather than the target behavior
- A single patch interpreted without replication
Note: The closest thing in this field to a controlled experiment.

# Your examples frame format louder than content
Kind: Tip
Tag: prompting
Published: 2025-07-18
Few-shot learning is pattern continuation. The model is extremely sensitive to surface form, sometimes more than to the actual reasoning you intended to demonstrate.
Make every example identical in structure, pick label words that are single clean tokens, and keep lengths similar. The more rigid the template, the more reliably the model fills in the next blank.
## Aha Moment
Few-shot examples frame surface form first. The model copies the pattern before it understands your intention.

## Try This
Make labels, punctuation, casing, separators, and example length boringly consistent.

## Watch For
- One example with a different label shape
- Multi-token labels competing with single-token labels
- Examples that demonstrate style drift instead of the task

# The logit lens: decode the middle layers
Kind: Method
Tag: interpretability
Published: 2025-07-05
The output projection is just a linear map, so nothing stops you from applying it to an earlier layer. The result is a rough readout of what the model would say if it had to stop thinking right now.
Often the correct answer appears in the middle layers and the later layers spend their budget refining phrasing and suppressing alternatives. Seeing that handoff demystifies a lot of model behavior.
## Aha Moment
The answer often appears before the wording is polished. Middle layers can show the thought taking shape.

## Try This
Decode intermediate layers and watch when the target answer first becomes linearly readable.

## Watch For
- Early guesses treated as final causal proof
- Layer readouts that shift after prompt paraphrases
- Ignoring later layers that suppress unsafe or irrelevant continuations