self-learn

Self-Learning Agent

3-phase autonomous agent: extract book content → validate against NotebookLM → extend beyond the source. Inspired by autoresearch — independent teams with information barriers.

All knowledge lives in Obsidian, never in this skill.

Configuration

obsidian_vault: "{vault}/knowledge/raw"  # Read {vault} from ${CLAUDE_PLUGIN_ROOT}/brain-os.config.md
notebooklm_bin: ~/.local/bin/notebooklm
scripts_dir: ${CLAUDE_PLUGIN_ROOT}/skills/self-learn/scripts

Phase 1: EXTRACT

Goal: Read the book → identify atomic concepts → write Obsidian notes.

1. Parse epub via scripts/lib/epub_parser.py to get chapter text
2. For EACH chapter sequentially: a. Identify atomic concepts — look for:
   • Named frameworks (e.g., "The Four Hats", "Thinking Time")
   • Core principles (e.g., "minimize dumb tax")
   • Processes/rituals (step-by-step methods)
   • Mental models (2nd-order consequences, checking assumptions)
   • Distinctions (operator vs owner, revenue vs profit)
   • Actionable tools (specific questions, decision frameworks) b. Create one note per concept (~100 words) using scripts/lib/note_writer.py c. Categorize into topic folders, link to related concepts from previous chapters
3. After all chapters: report total topics extracted and category breakdown

Knowledge Path

{obsidian_vault}/{book-slug}/
  ├── decision-making/
  ├── business-operations/
  ├── leadership/
  └── [discovered categories]/

Note Template

---
source: "{book_title}"
author: "{author}"
chapter: {chapter_number}
tags: [{tags}]
---

# {Concept Title}

{~100 words: concise, crystal clear, key point only}

## Key Insight
{One sentence takeaway}

## Related
- [[related-concept-1]]
- [[related-concept-2]]

Phase 2: VALIDATE (Autoresearch Loop)

Goal: Compare agent knowledge against NotebookLM. Patch gaps until mastery.

Architecture — 4 independent roles with information barriers

1. Question Generator — creates questions without seeing answers
2. Knowledge Agent — answers ONLY from Obsidian notes (never the book directly)
3. Oracle (NotebookLM) — answers from full book via notebooklm ask
4. Judge — scores similarity without knowing which answer is agent vs oracle

NotebookLM Setup

~/.local/bin/notebooklm use {notebook_id}    # Set active notebook FIRST
~/.local/bin/notebooklm ask "{question}"      # Then ask questions

Process

1. Generate questions: total = max(100, num_topics × 3)
   • 50% topic-based (from Obsidian note content) via scripts/lib/question_gen.py
   • 30% blind (from chapter titles, cross-cutting)
   • 20% adversarial (designed to expose shallow understanding)
2. For EACH question: a. Agent answers by reading relevant Obsidian notes only b. NotebookLM answers via notebooklm ask "{question}" c. Judge scores using scripts/lib/judge.py prompt (0-100 scale) d. If score < 95: identify what's missing → update/create Obsidian notes
3. Track scores in validation log
4. Stop when: 100% pass rate — ALL questions score ≥ 95
5. If not passing: generate NEW questions targeting weak topics, repeat

Wrong concepts are catastrophic. Fix conceptual errors before missing details. A note that says the wrong thing is far worse than a note missing a detail.

Validation Log

Round 1: 85/100 passed (85%) — weak: [business-operations, leadership-hats]
Round 2: 94/100 passed (94%) — weak: [accountability-structures]
Round 3: 100/100 passed (100%) — MASTERY ACHIEVED

Phase 3: EXTEND + INGEST

3a. Create Structured Book Note

After validation passes, create the structured book note directly (previously a separate /ingest step):

1. Read all validated atomic notes from {obsidian_vault}/{book-slug}/
2. Synthesize into a single structured book note at {vault}/knowledge/books/{book-slug}.md
3. Use the Book Note Template (see /ingest skill for template)
4. Extract: metadata, key concepts, frameworks, actionable insights, quotes, vault connections
5. Mark ingested: true in _validation/audit-flag.json

3b. Extend Beyond Source

1. Synthesize with adjacent thinkers — tagged source: "synthesis"
2. Apply to new domains — in applied/{domain}/, tagged source: "application"
3. Research current developments — tagged source: "research-{date}"

Extension notes use same template but different source tags.

After Completion

Pipeline continues autonomously: /verify → /absorb → commit + push → notify. Phase 2 must reach 100% pass at ≥95 before Phase 3 runs. See /study for full pipeline.

Usage

/self-learn --book <epub_path> --notebook-id <nlm_id>
/self-learn --phase extract --book <epub_path>
/self-learn --phase validate --notebook-id <nlm_id>
/self-learn --phase extend

Status — always run the script, never generate manually

python3 ${CLAUDE_PLUGIN_ROOT}/skills/self-learn/scripts/summary.py {vault}/knowledge/raw/the-road-less-stupid --status
python3 ${CLAUDE_PLUGIN_ROOT}/skills/self-learn/scripts/summary.py {vault}/knowledge/raw/the-road-less-stupid --phase3
python3 ${CLAUDE_PLUGIN_ROOT}/skills/self-learn/scripts/summary.py {vault}/knowledge/raw/the-road-less-stupid

Available Books

Book	Vault Path	NotebookLM ID
The Road Less Stupid	knowledge/raw/the-road-less-stupid	9c6d742c-d512-42f6-a42c-ccdd993ef19f
Start with Why	knowledge/raw/start-with-why	e990cedd-7dca-4812-824e-7a9e757f0065
Changing World Order	knowledge/raw/dalio-changing-world-order	bf8f46c4-0dfb-4f5e-9013-7a9e757f0065

Outcome log

Follow skill-spec.md § 11. Append to {vault}/daily/skill-outcomes/self-learn.log:

{date} | self-learn | {action} | ~/work/brain-os-plugin | knowledge/raw/{slug}/ | commit:{hash} | {result}

• action: extract, validate, or extend
• result: pass if phase completed (validate = 100% pass rate), partial if progress but not complete, fail if stuck or errors
• Optional: args="--book {epub} --notebook-id {id}", score={pass_rate}%