For agents: start with the agent guide. Every page is Markdown by default; add Accept: text/html for HTML.

Learn during real work

Teach the technology and principles in play. Use the project as a worked example. Keep the current task moving.¶

Start and stop

• Start only when the operator enables learn mode¹⁰.¶
• When the operator turns learn mode on, act at once. Read learnPreferences, call learnNext with no arguments, and present the returned concept.¶
• If the tools are not ready, say in one line how to ask: “Say next concept any time, or keep working and I will explain ideas as they appear.”¶
• Keep questions off unless the operator explicitly enables them.¶
• Stop tool calls and explanations when the operator turns learn mode off.¶
• Keep existing progress when stopping. Turning the mode off does not erase history.¶
• Record dismissed when the operator asks to stop presenting one concept.¶
• Do not promise a reset or signed-in sync. Those tools are not available.¶

Use the local tools

The local tools need a rat-stack checkout whose CLI has the learn commands. Check before you build or pull:¶

git -C <checkout> status --porcelain
node <checkout>/apps/cli/dist/cli.js --help

• The help output must list learnNext, learnRecord, learnPreferences, and learnSetPreferences.¶
• Never pull, reset, or stash in a checkout with local changes. Clone a separate copy for the learn tools instead.¶
• Do not use the operator’s working checkout when it is behind or dirty. A fresh clone keeps their work untouched.¶

Clone, then build the CLI:¶

git clone --depth 1 https://github.com/joelhooks/rat-stack.git rat-stack-learn
cd rat-stack-learn
pnpm install
pnpm turbo run build --filter=@rat-stack/cli

Configure the agent’s stdio MCP client to run this command from that checkout:¶

node apps/cli/dist/cli.js mcp

Add --devtools when checking recorded calls. The local tool descriptions say they use the local event log. Public MCP²⁷ and HTTP progress tools use supplied progress and store no progress.¶

Feedback is a separate signed-in operation. It does not synchronize learning progress.¶

The default local store is ~/.rat-learn/. RAT_LEARN_DIRECTORY selects another local directory. RAT_LEARN_ORIGIN selects the public deck origin.¶

Display preferences

~/.rat-learn/preferences.json holds three display choices. It stays on the machine.¶

• Read it with learnPreferences. A missing file means the defaults.¶
• Change it with learnSetPreferences only when the operator asks.¶
• narrow: keep the snippet and diagram one after the other, as plain text in the reply.¶
• wide: the diagram may sit beside the snippet when the display allows it.¶
• rich: the agent’s own display may render the diagram, but uses only the card’s fields.¶
• terse: show the service declaration and the line that uses or provides it. full: show the whole snippet.¶

Select and present

1. Read learnDeck and match public terms to the ideas in the work.¶
2. Keep that matching step local. Select only real concept ids from the deck.¶
3. Call local learnNext when an Effect, Alchemy, or XState idea appears in the work.¶
4. Set context.ids to those concept ids and context.at to the current Unix time in milliseconds.¶
5. You may use context.terms instead of ids. Pass only public terms from the deck.¶
6. The tool resolves a term’s owner and offers an unmet prerequisite first.¶
7. Omit context to follow the ordered learning path. Dismissed prerequisites do not block it.¶
8. Set context.asked only when the operator asks for an explanation.¶
9. Omit progress. Omitted progress means none yet, and the local tools read their own log.¶
10. Follow the returned depth. Present one relevant concept at a time.¶

Use the card’s snippet and diagram exactly as served. The content build compiles every snippet against the pinned versions. Never write your own snippet or diagram for a card.¶

A card without a snippet uses its page’s pinned excerpt and immutable revision. Never substitute current source for the pin. State when that card has no suitable excerpt.¶

Call local learnRecord only after presenting. Use a shown event with the concept id, actual depth, current timestamp, and version 1. Selection alone does not record a presentation.¶

Record what happened

• Record used when the operator applies the idea in their own code.¶
• Agent-written code alone does not establish operator use.¶
• Record got-it or skipped only from the operator’s words.¶
• Do not infer understanding from silence, elapsed time, or reading a page.¶
• Never send the observed code or the operator’s words with the event.¶
• If recording fails, say that progress was not confirmed. Preserve the log and follow the typed error.¶

Keep it quiet

• Never quiz unless the operator opts in to questions.¶
• With questions enabled, ask at most one short check-in per long session.¶
• Use a due concept for that check-in. Accept skipping it immediately.¶
• Do not ask permission again after the operator has enabled the mode.¶
• Stop check-ins when the operator disables questions.¶
• Do not introduce unrelated concepts to fill a learning quota.¶

Without MCP

Read the public deck through either surface²³:¶

curl -fsS https://ratstack.sh/_content/learn.json
curl -fsS https://ratstack.sh/api/learnDeck \
  -H 'content-type: application/json' --data '{}'

Use the CLI commands for selection and recording. They use the same local store as stdio MCP. Read the system check¹⁰ for a complete round-trip.¶

Build first, then run node apps/cli/dist/cli.js learnNext. Add a --context JSON flag for concepts in play. Record with learnRecord --event. Use current timestamps for real learning events.¶

If the deck or CLI is unavailable, report that limit and continue the task. Do not claim that progress was saved. Never hand-edit the JSONL files.¶

Leave feedback with phone approval

Use this flow only when the operator asks to leave feedback on a public card. The server exposes these tools only when AUTH_ENABLED is true. If they are absent from MCP, report that feedback sign-in is unavailable on that surface. Do not retry.¶

1. Draft the feedback with the operator. Keep it about the public card.¶
2. Call learnFeedbackStart with an empty input.¶
3. Show the returned userCode and verificationUrl to the operator.¶
4. The operator opens the URL on their phone, signs in, and checks the code.¶
5. The operator chooses Approve feedback only or Deny.¶
6. Call learnFeedbackPoll with the returned deviceCode. Wait at least the returned interval between calls.¶
7. Stop on denied or expired. On polled-too-fast, wait the returned interval before another poll.¶
8. After approved, call learnFeedback with token, cardId, and the agreed feedback text.¶
9. Confirm storage only after the capability²¹ returns a feedback id.¶

The credential permits learnFeedback only. It is not a sign-in session or a credential for another tool. It expires with the device grant. The provider’s device-code lifetime is thirty minutes, and its polling interval is five seconds. A new device flow is required after expiry. Do not use the stock /auth/device/token endpoint; it is disabled.¶

Keep the credential out of local learning JSONL, feedback text, examples, and shared messages. Do not include private code, prompts, paths, credentials, or names in feedback. A failed call does not prove storage. Automatic retry can store a second feedback row; ask before repeating an uncertain write.¶

Offer a skill in the session

Learn mode suggests skills. It never installs them.¶

1. When a card’s route is /skills/<name>, or the work matches a skill, name the skill in one line and link https://ratstack.sh/skills/<name>.¶
2. Ask before loading it. Load only what the operator wants.¶
3. To load it, fetch https://ratstack.sh/.well-known/agent-skills/<name>/SKILL.md and follow it for the rest of this session.¶
4. Never run a skills installer. Never write a skill into a global, user or project skills folder.¶
5. If the operator wants a skill kept for good, they choose where it goes.¶

Privacy

Keep progress local. Preferences stay local too.¶

Use the public endpoints for deck and card reads. Do not send progress to public selection or recording endpoints by default.¶

Requests contain concept ids, timestamps, depth, familiarity, and dismissal state only. Never include code, file paths, names, repository names, or prompts. The event log is authoritative. The CLI rebuilds cards.jsonl and tasks.jsonl from it.¶

Last changed 2026-10-10 in 237eb0c. Source on GitHub. Change log.