learning

A study system that uses AI (Claude Code) as a guide to end shallow studying. Every session follows a prescription built from your study habits (me/prescription.md).

Getting started

1. Clone this repo and open it with Claude Code. On first open you may be asked to approve the project hooks (used for session logging).
2. Create me/ yourself (prescription, curriculum, plan, config). Follow the format of the existing files.
3. /study: picks today's item from the plan and starts. /study <topic> studies an off-plan topic.

Structure

Path	Contents	git
.claude/skills/	study (runs the session), git-commit	public
.claude/agents/	researcher (web research), svg-maker (spatial/geometric figures)	public
.claude/tools/	plan.py (next item, progress), stats.py (review schedule, confidence stats)	public
.claude/hooks/	md_log.py (live session logging)	public
notes/<stage>/<concept>.md	Organized concept notes	public
me/	prescription.md, curriculum.md, plan.tsv, config.json	local
log/	quiz.tsv, sessions.tsv, sessions/ (raw conversations)	local

Methodology

Taken from three sources.

Source	What was taken
DrCatHicks/learning-opportunities	Learning-science grounding. Predict first, say plainly when an answer is wrong, recall check at session start
amosblomqvist/learn	plan → teach. Concept dependency DAG, "how could I have discovered this?", live session logging (md-log)
rlaope/estudy	Don't polish notes, just accumulate them. Book → LLM DFS → on-demand project

Principles every prescription rests on:

• Generation and testing effects: producing and retrieving an answer sticks longer than reading or watching. So sessions center on quizzes and teach-back (3-5 sentences in your own words).
• No pretest: research says guessing before the material helps, but for this user it only causes frustration, so quizzes come after teaching. The order is goal/level check → material → check quiz.
• Spaced repetition: review is the entry fee to a session. Each correct answer moves the next review out 1 → 3 → 7 → 14 → 30 → 60 days; a wrong answer resets to 1 day.
• Illusion of fluency: easy reading feels like understanding. Don't re-read the material right before a quiz. Re-read only after a miss, and only the part you missed.
• Clear correction: when wrong, say so immediately. Vague correction doesn't work.
• Calibration: record a confidence level on every quiz to see in numbers whether you "know but doubt" or "don't know but feel sure".
• On-demand: put the endpoint of studying at a concrete point in the project (or interview) you are building.

Session flow

There is no time limit. A session ends when the item's scope is finished or you stop.

1. Review quiz → goal and level check (no quiz)
2. Plan: approve the concept dependency DAG, nodes you already know are marked check-only
3. Material (video or textbook, links provided)
4. Teach loop: a check quiz per node; only missed nodes get explained
5. 3-5 sentences in your own words → organize the concept note → record accuracy with plan.py log → commit

During /study the conversation is logged live to log/sessions/. Open that folder as an Obsidian vault and read the file alongside. Obsidian is read-only here: no plugins or themes, and plain markdown links instead of [[wikilinks]] (Settings → Files and links → turn off "Use [[Wikilinks]]").