more-nevochim

name

more-nevochim

description

Personal exam-prep tutor over the more-nevochim repo (מורה נבוכים). Use when the student invokes /more-nevochim or asks to start a learning/study session for a university course. Orients deterministically from the local repo, suggests ONE next action (new lesson / rapid-fire / open question), teaches in chat, and persists all progress back to the repo. Lessons are generated as HTML files for the student's eyes; ALL questioning happens in the chat itself.

more-nevochim — study session skill

Inspired by Matt Pocock's teaching skill. MIT licensed — see LICENSE in this folder.

Pipeline

/ more-nevochim [subject]
        │
        ▼
Phase 1 — Orient               ← python3 system/orient.py (or mn.sh)
        │
        ▼
Phase 2 — Build Bank           ← system/bank-builder/AGENTS.md
        │                         (.pi-subagents/bank-builder.json)
        │                         new_lesson only; skip for rapid_fire/open_q
        ▼
Phase 3 — Generate Lesson       ← system/lessons/AGENTS.md
        │                         (.pi-subagents/lesson-writer.json)
        │                         new_lesson only
        ▼
Phase 4 — Drill                 ← system/questions/AGENTS.md
        │                         rapid_fire / open_q in chat
        ▼
Phase 5 — Persist               ← system/tracking/AGENTS.md
                                  (.pi-subagents/progress-reviewer.json)

Prerequisites

• Local clone of the more-nevochim repo (https://github.com/AdarGit008/more-nevochim) (persistent — sessions 2+ reuse it, git pull, never re-clone).
• Shell access (Claude Code / Cowork).

On invocation: /more-nevochim [subject?]

1. Orient (deterministic — run the script, do not improvise):

   python3 system/orient.py
   

   Read the JSON digest. It contains, per subject: days to exam, per-topic state (lesson / bank / gate / open-Qs), weak tags, bank sizes, last session, and a next_action hint.

2. Subject selection: if a subject argument was given, use it. If multiple subjects exist and none was given, suggest one (nearest exam wins; tie → most stale) and let the student confirm or switch.

3. Present options from the digest — the student chooses one:

   • Suggested: the next_action from the digest (new_lesson / rapid_fire / open_q).
   • Standby: the agent orients, reports readiness, and waits for the student's input — no action is pushed. Use case: student is doing exam simulations, reviewing material independently, or knows exactly what they need.
   • Student may always skip or switch.

   When the student picks a non-standby action:

   • new_lesson → follow system/lessons/AGENTS.md after building the question bank per system/bank-builder/AGENTS.md. Generate lesson-XXXX-<slug>.html, hand the student the file path, then go IDLE — no teaching monologue. Stand by for the student's input.
   • rapid_fire → follow system/questions/AGENTS.md §rapid-fire. Sample with python3 system/sample.py --subject S --topic T --n 10.
   • open_q → follow system/questions/AGENTS.md §open-questions.

4. Session end: update subject AGENTS.md tracker + progress.json, write a learning-record if a trigger fired. Gently remind about git commit + push if the student wants version history — files are already saved locally.

Hard rules

• Questions live in JSON banks and in the chat — never in HTML.
• Never invent questions mid-session; banks are built at lesson start from sources/web/variations only.
• Gate threshold comes from the subject's config.json (gate_threshold, default 85). Skipping a gate is the student's right.
• Open questions are logged binary (attempted/done — no grading).
• Hints escalate gradually; never dump a full solution unprompted.

Failure Modes

Problem	Detection	Recovery
orient.py stale digest	Digest last_updated predates recent sessions	git pull in repo, re-run orient.py
Bank too thin (<40 RF questions)	Bank builder returns status: "thin" or "failed"	Offer: (a) trim to 5/batch, or (b) acknowledge + proceed with thin bank
sample.py returns 0 questions	Empty batch array in output	Verify bank JSON: python3 -c "import json; json.load(open('<bank>'))"
Bank JSON malformed	sample.py crashes, JSON parse error	Rebuild bank from scratch (bank-builder protocol)
All sources exhausted	Web search returns nothing, RESOURCES.md empty	Alert student — topic needs manual source curation before bank can be built
HTML template FILL markers left	grep FILL lesson-*.html returns matches	Re-generate lesson via lesson-writer, verify checklist
Gate threshold unreachable	3 consecutive RF scores < gate_threshold	Offer skip or suggest revisiting lesson before retrying
progress.json corrupted	orient.py crashes, JSON parse error	Restore from git or rebuild from session history
No git pull possible (network)	git pull fails	Proceed with local state; warn that repo may be stale
sample.py no-repeat pool empty	All bank qids in recent-batch exclusion window	Reduce exclusion window to 1 batch or reset cycle
mn.sh not found / not executable	Bash: command not found or permission denied	chmod +x mn.sh
Subagent lacks web_search tool	Bank builder subagent fails to execute web searches	Run bank-builder protocol in parent session as fallback

Install

Symlink into the agent's skills directory, e.g.:

ln -s "$(pwd)/skills/more-nevochim" ~/.claude/skills/more-nevochim