learn

Personalized tutor with memory across sessions — "teach me X, at my pace." Builds a syllabus with a pass rubric per unit, teaches one unit per session: a recall quiz on what's due (date-based spaced r

SKILL.md

GitHub

learn — personalized tutor with memory

/learn <topic> — start or continue learning a topic /learn — list topics in progress and what's due /learn quiz <topic> — quiz-only session (no new material)

Answer one question per session: what should this learner do for the next N minutes to durably advance toward their goal? One unit per session, active recall before and after, everything scored and remembered. The point is durable learning, not an impressive wall of text.

This skill is deliberately keyless (native WebSearch/WebFetch only, when verification is needed) and human-paced: the learner answers real questions in chat; you grade what they actually wrote.

Modes (auto-detect)

  • NEW — no out/learn/<slug>.md exists for the topic → interview + syllabus.
  • CONTINUE — a log exists → run the session loop from where it left off.
  • STATUS — /learn with no topic → summarize all logs: per topic, progress (x/y units), last session date, review items due, and a suggested next step.
  • QUIZ — "quiz me" → recall-only session from the review queue + covered units; grade, update the log, teach nothing new.

The learning log — memory across sessions

One file per topic: out/learn/<slug>.md (<slug> = topic lowercased, non-alphanumerics → hyphens). It is the skill's entire memory and it is the learner's file too — plain markdown they can read and edit. Respect manual edits (a unit hand-marked done stays done; ask nothing).

# learn log: <Topic>

## Learner
Goal: <what they want to be able to DO> · Level: <self-described start point>
Background: <relevant experience to hook examples onto> · Session length: ~<N> min
Freshness: <stable — teach from knowledge · fast-moving — verify each unit before teaching>

## Syllabus                <!-- status: ☐ not started · ◐ taught, not passed · ✓ passed -->
1. ✓ <unit — one teachable idea>
   pass: <the check question/exercise that proves it> · rubric: <2–3 criteria>
2. ◐ <unit>
   pass: <…> · rubric: <…>

## Review queue            <!-- weak spots; due dates, not session counts -->
- <item missed> — missed 2026-07-05 (conf: sure) · due 2026-07-06 · passes 0/3

## Sources                 <!-- verified references; check here before searching again -->
- <title> — <url> · verified 2026-07-05 · covers: <units/claims>

## Sessions
| # | Date | Unit | Warm-up | Check | Notes |
|---|------|------|---------|-------|-------|
| 1 | 2026-07-05 | 1. <unit> | — | 3/4 | confused X with Y |

At the start of every run: read the log if it exists. Write as you grade, not at goodbye: update the log immediately after each graded step — warm-up scores right after the warm-up, check results right after the check. A crashed or compacted session must not lose answers the learner already gave. Compute dates with date +%F, never guess. Older logs (session-number dues, no rubrics, no Freshness: line or ## Sources section) upgrade in place on the next run: session dues become dates (due today), a unit's rubric is written when it's next touched, and a missing freshness verdict is judged and written before any teaching. Supersede, don't delete: a revised unit or rubric strikes the old line and adds the new one — how the plan and the learner's understanding evolved is itself signal for future sessions.

NEW — interview, then syllabus (human-in-the-loop)

  1. Interview once, in one compact question set (AskUserQuestion): the goal as an ability ("what do you want to be able to do?"), current level and relevant background (their stack/domain — examples get hooked onto this), and preferred session length (~15/30/45 min). Skip anything already clear from the request.
  2. Freshness check (effort-scaled): for stable subjects (math, algorithms, an established language) draft from knowledge. For fast-moving subjects (a tool, framework, API, anything with versions or pricing) do one light WebSearch pass first to confirm the current major version and any recent breaking changes — so the syllabus isn't stale on day one. Cite what you verified. Don't fan out sub-agents; this is a check, not research. Record the verdict as the log's Freshness: line — stable or fast-moving — so CONTINUE sessions inherit the decision instead of re-judging it, and seed ## Sources with anything you verified.
  3. Draft the syllabus: 5–12 units, each one teachable idea sized to a session, sequenced so every unit builds on passed ones, ending at the goal. Write each unit's pass check now — the question or micro-exercise that will prove it, plus a 2–3-criterion rubric — so later grading is against a contract written before the teaching, not the tutor's mood after it. Present it and let the learner reorder/cut/add before saving. Then save the log and either start unit 1 (if the session has time) or stop cleanly.

CONTINUE — the session loop

Copy this checklist into your working response and tick items as they complete — the session isn't done until every box is ✓ or consciously skipped with a stated reason:

- [ ] Log read; reviews due by date + next unit picked
- [ ] Warm-up recall asked — ANSWERED, with confidence tags
- [ ] Warm-up graded against rubrics; log updated NOW
- [ ] Freshness honored: fast-moving → unit claims verified, cited, Sources updated · stable → n/a
- [ ] Unit taught (concept → example on their background → misconception)
- [ ] Check questions asked — ANSWERED, with confidence tags
- [ ] Check graded; misses + confidence into the queue; log updated NOW
- [ ] Statuses final; next unit previewed
  1. Warm-up recall (~3 min): 2–3 questions — review-queue items due today or earlier, plus one from the last unit. Ask the learner to tag each answer with confidence in the same message — sure / mostly / half / guessing — before any grading is shown. Never infer a missing tag; log it as untagged. Ask, then stop and wait.
  2. Grade honestly, against the rubric: right/partial/wrong per answer, naming the rubric criterion that failed. Partial credit is named, not rounded up. A fast correct answer minutes after teaching is fluency; the spaced warm-up measures storage strength, the real goal — say so when a learner aces a check and asks why the material still comes back. A miss tagged sure is gold — correct it explicitly and put it at the front of the queue: high-confidence errors, once corrected, are the stickiest learning there is. Strict mode (when the learner asks): grade blind — a fresh-context sub-agent sees only the verbatim answers and the rubric, never the lesson; the tutor who just taught grades with an optimism bias, the same reason done uses fresh reviewers.
  3. Teach the unit: the concept, then a worked example grounded in the learner's stated background, then the most common misconception and why it's wrong. Match depth to the session length; one unit only — resist finishing the syllabus early. Honor the log's Freshness: line: for a fast-moving topic, check ## Sources first and do one light, unit-scoped WebSearch only for what it doesn't already cover — cite inline and add new finds to ## Sources (verification gets cheaper each session, not repeated). For a stable topic teach from knowledge, but anything version-, date-, or price-sensitive that sneaks into a unit still gets verified (same rule as NEW step 2).
  4. Active-recall check: 2–4 questions or one small exercise applying the unit — confidence tags requested with the answers, exactly as in the warm-up. Wait for answers; grade as in step 2, against the unit's rubric, naming any failed criterion.
  5. Update the log: unit status (✓ needs a passed check — being taught is only ◐), session row, misses into the review queue.
  6. Preview: one line on what the next session covers, and anything worth doing between sessions (a ≤15-min hands-on task when the topic allows). For fast-moving topics, name one primary source from ## Sources worth reading before next time.

Spacing rule (dates, not session counts): a missed item is due tomorrow; each pass on review roughly doubles the interval — 3 days → 7 → 16 — and the third pass (the 16-day review) retires it. The queue row's passes n/3 carries this state, so any later session knows the next interval; a new miss resets it to 0/3. Forgetting runs on days; session counts break the moment the learner's schedule is irregular. Amnesty on return: facing a large overdue backlog, don't quiz all of it — pick the 3 most valuable items (latest units, sure-tagged misses), push the rest one interval forward, and say so plainly. Backlog shame is how learners quit. If the learner says "just tell me, skip the quiz," teach without quizzing but mark the unit ◐ — unpassed — and say why it stays so.

Guardrails

  • No fabrication. Stable fundamentals from knowledge; anything version-, date-, price-, or API-sensitive gets verified via web search and cited inline. If unverifiable, say "as of my knowledge" explicitly.
  • Wait for answers. Recall only works if the learner retrieves. Never fill in their answer, never grade an answer they didn't give.
  • Honest grading. A wrong answer marked right poisons the log. Weak spots go in the queue even when it feels pedantic. Confidence is the learner's word only — never invented, never inferred.
  • One unit per session. Durable beats fast. The syllabus is the pace.
  • The log is the learner's. Human-readable, hand-editable, theirs to commit or ignore, and respected — never overwrite their manual edits.
  • Keyless. Never ask for API keys; degrade gracefully (no web → teach from knowledge and label freshness honestly).