Learn by doing

Learn By Doing

Speak to the user in their language; these instructions are in English for the agent.

A generative skill: produces a markdown file under learn/ containing 3–7 graded exercises plus a hidden solutions section. The user attempts the exercises; the agent verifies each answer against the real code/docs when the user submits them. The artefact (the exercises file) persists in the project and can be reopened later — see ADR 0002 for why we don't schedule those reopenings.

Quick start

User: /learn-by-doing src/orders/payment.ts
Agent: Project read first. Stack: TypeScript + Express 4.21, ESM modules.
       I'll generate 5 exercises of ascending difficulty, save them to
       learn/payment-exercises.md, and hide solutions behind a fold.
       Bring me each answer when you're ready — I'll verify against payment.ts.

After generation, the user works through the file, brings answers back into the conversation, and the agent verifies.

Workflows

Mode detection

Inspect the argument:

• Code-mode — argument is a file, function, or PR. Exercises target this code (refactor, extend, recreate from spec, find the bug).
• Topic-mode — argument is a topic. Exercises target applying the concept (implement, design, choose between alternatives) within the project's actual stack.

Session checklist (copy into your reply and tick as you go)

Generation:
- [ ] Project read complete (stack + versions identified)
- [ ] Artefact read (code-mode) or docs fetched (topic-mode)
- [ ] 3–7 exercises drafted with ascending difficulty (E1 → E7 ladder)
- [ ] E4 (Debug) included when code-mode
- [ ] Every solution traces to file:line or cited doc
- [ ] Solutions folded under <details> tag (none above the fold)
- [ ] File written to learn/<slug>-exercises.md
- [ ] Filename + path confirmed with user before write if collision

Per-answer verification (each time the user brings an answer):
- [ ] User's answer read carefully (not skimmed)
- [ ] Verified against artefact with file:line or doc citation
- [ ] Marked pass / partial / fail with evidence
- [ ] On fail: one targeted question + line reference + retry (no spoiler)
- [ ] On pass: user decides whether to proceed to next exercise

Generation protocol

Read EXERCISES.md for the full generation reference (exercise ladder E1–E7 + file format). Summary of the steps:

1. Project read — identify stack + versions from manifest files; never use a generic version.
2. Read artefact / fetch docs — every solution must trace to a citation.
3. Pick 3–7 exercises from the ladder (E1 Reproduce → E7 Teach). Include E4 (Debug) in code-mode.
4. Write the file to learn/<slug>-exercises.md with solutions folded under <details>.

Verification protocol (when the user brings an answer)

1. Read the user's answer. Do not skim — read it like a code review.
2. Verify against the artefact (Read tool for code, WebFetch / context7 for docs). Mark pass / partial / fail with the specific divergence.
3. Cite. Every claim about "the correct answer" carries a file:line or a doc link.
4. If failed, do not reveal the solution. Ask one targeted question and point at the relevant artefact line(s). Let the user retry.
5. If passed, move to the next exercise (the user decides when).

Workspace mode

If CURRICULUM.md exists in the current directory, the skill is running inside a Scenario B workspace. In workspace mode:

• Exercises go to exercises/<slug>-exercises.md (not learn/).
• Read MISSION.md before generating: use the stated goal to shape exercise difficulty and context. A user learning chess for casual play needs different exercises than one studying openings competitively.
• In topic-mode (Scenario B, no existing code): read CURRICULUM.md to identify which lesson the exercises should reinforce, then target that lesson's concepts.

If CURRICULUM.md is absent (legacy mode), use learn/ paths as before.

File hygiene

• One file per session (don't merge into an existing file — the user might want different sets for different goals).
• File names: <slug>-exercises.md, slug derived from topic or source file.
• In legacy mode: learn/<slug>-exercises.md. In workspace mode: exercises/<slug>-exercises.md.

Anti-patterns

• Don't put solutions above the fold, even partially — visible solutions defeat the entire skill, because the act that produces learning is the user attempting before consulting; one peek collapses recall into recognition.
• Don't write solutions the agent invented from memory — every solution must trace to the artefact, otherwise the user gets verified against the agent's confidence rather than the real source; that's borrowed confidence at the verification layer.
• Don't write generic exercises ("implement quicksort") — exercises must use this stack at this version on this problem, otherwise the transfer doesn't connect to the work the user actually does.
• Don't reveal the answer when the user is stuck — the productive struggle is the learning signal; offer a targeted question + line reference + retry, and let the user re-attempt with the source closed again.
• Don't generate more than 7 exercises in one file — past 7, fatigue dominates and exercise quality drops; the user answers shallower and learns less per exercise.
• Don't skip E4 (debug) in code-mode — debug exercises produce the highest transfer signal because they require a working mental model to even identify the broken assumption.

Governing principles (this skill satisfies all five)

1. Agent withholds — solutions folded; verification asks before telling.
2. Student speaks first — the user produces every answer.
3. Artefact is the judge — every solution and every verification is grounded in real code/docs.
4. Source fidelity — stack version read before generation; every solution carries a citation to an authoritative source. Policy: ../../docs/sources.md.
5. Exit is a transfer test — the last exercises (E5–E7) are transfer-tests by construction.