| name | study |
| description | Start a study session — the main entry point for learning. Shows progress, runs due spaced-repetition reviews, grades free-text answers, updates SM-2 and Obsidian. Cost-aware: recommends a cheaper model and delegates grading to Haiku. Use when: 'study', 'study session', 'lass uns lernen', 'review time', 'let's learn', 'practice'. |
Study — Your Learning Session
The umbrella entry point. One command instead of juggling learning-status + quiz + check-answer separately.
Cost Awareness (do this first)
Study sessions don't need a frontier model — questions come deterministically from JSONL, and answers are short. First action of every session: check which model is running. If it's an Opus-class model, tell the user once:
Tip: study sessions run fine on a cheaper model — /model haiku (or sonnet) saves budget. Want to switch first, or continue?
Don't block on it — if they say continue (or ignore it), proceed. Never repeat the tip within a session.
For free-text answer grading (the only step that needs real judgment), delegate to a cheap subagent regardless of session model: spawn an Agent with model: "haiku" whose prompt contains the question, the correct answer (from the pattern's explanation field), and the user's answer — it returns a 1-5 score plus one sentence on what was missed. Batch multiple answers into one agent call when possible.
Session Flow
- Status snapshot (3-4 lines max): streak, due count, weakest category. Read from
~/.learn-while-code/data/concepts.jsonl and quiz-history.jsonl. No giant dashboard — this is a warm-up, not a report. (Full dashboard stays in learning-status.)
- Pick up to 5 due concepts — same rules as quiz:
encountered first, then nextReview <= today, interleave categories, prioritize cross-project frequency.
- For each concept, ask the
whyQuestion and let the user choose how to answer:
- Self-rate: they think, you reveal the answer, they rate 1-5
- Free-text: they write their answer, you grade via the Haiku subagent (score 1-5 + one-line feedback), show the correct answer
- Update after each answer: SM-2 via
sm2Update(), append to quiz-history.jsonl
- End of session: summary table (concept, score, new status, next review), then sync changed concepts + indexes to Obsidian if configured
- Close with one pointer: the single weakest area and how to hit it ("most of your SAP concepts are still unseen — code in MCP-CX-Operations and they'll get detected")
Relationship to Other Skills
| Skill | Keep using it for |
|---|
study | The default — regular learning sessions |
quiz | Direct quiz without status/warm-up (study calls the same logic) |
learning-status | Full dashboard with all bars and tables |
check-answer | Grading a single answer outside a session (e.g. from an Obsidian note) |
learn-project | Writing project knowledge notes — documentation, not review |
Gotchas
- Max 5 questions; quiz fatigue kills retention.
- The status snapshot is 3-4 lines. Resist the dashboard urge.
- Free-text grading is generous: core WHY right = 4+. Edge cases are bonus, not requirement.
- If nothing is due: say so, show the next review date, suggest
learn-project for the current repo instead.
- Match the user's language.
