| name | learn |
| description | Generate and deliver a learning exercise based on what the user is forgetting. Orchestrates different learning techniques. |
Learn — The Intervention Engine
This is the master learning skill. It decides which technique to use and executes it.
Steps
-
Get the data — call predict endpoints (recommendations, learner state, strategy signature, calibration)
-
Decide the technique based on the API response:
| Condition | Technique | Why |
|---|
| Lapses > 0 for this concept | Spaced Retrieval | They've forgotten before — need active recall |
| Recommendation action = "quiz" | Spaced Retrieval | API recommends quiz |
| Strategy = "overconfident" | Calibration Check | They think they know it but don't |
| Strategy = "cram_dependent" | Interleaving | Break the cramming pattern |
| Multiple related concepts decaying | Interleaving | Practice related concepts together |
| Concept has high stability but hasn't been tested | Delayed Probe | Quick surprise check |
| Recommendation action = "review" | Replay Consolidation | Multi-phase memory replay |
| User asks to explain something | Teach Back | Explain it to prove mastery |
- Check the stability budget before intervening:
curl -s -H "x-api-key: $TOKEN" \
"https://cognition-api.fly.dev/v1/stability/budget?user_id=USER_ID"
If the budget says the intervention would destabilize learning, defer or soften the exercise.
-
Execute the technique — load the specific technique skill and follow its instructions.
-
Record results via the reflect flow — POST quiz_answer or delayed_probe events back to the API.
Delivery Modes
Terminal mode (default): Present exercises as interactive markdown. Ask questions, wait for answers, give feedback.
Rich UI mode: When the exercise benefits from visual interaction (knowledge maps, multi-question quizzes), generate an HTML file using the scripts in scripts/ and open it in the browser:
python ${CLAUDE_SKILL_DIR}/../scripts/quiz.py --output /tmp/cognition-quiz.html
After the Exercise
Always show:
- How many questions right/wrong
- How stability changed for each concept
- When the next review is predicted
- Encouragement based on progress
