repo-history-book

name	repo-history-book
description	Build a durable, evidence-backed "book" of how an engineering project evolved by mining git history, PRs, releases, and docs. Use when the user wants repository archaeology, commit-by-commit walkthroughs, engineering learning extraction, pivot analysis, or a readable chronology of what the team changed, learned, reverted, and refined over time.
hooks	{"PostToolUse":[{"matcher":"*","hooks":[{"type":"command","command":"python3 \"$HOME/.claude/skills/skill-feedback/scripts/skill-event.py\" --skill repo-history-book --event skill_activated --agent-harness claude-code --quiet","timeout":5}]}]}

Repo History Book

Use this skill when the user wants more than a changelog:

• "walk me through how this repo evolved"
• "what did the engineers learn building this?"
• "turn the git history into a readable book"
• "go commit by commit / PR by PR"
• "extract pivots, aha moments, pain points, and refactors"

The core problem is memory. Do not rely on the model remembering hundreds of commits. Externalize state early and keep it hierarchical.

Outcome

Produce a repository history "book" with:

1. Exhaustive evidence store — all commits, grouped into stable chunks
2. Durable intermediate notes — commit notes, day notes, phase notes, pivot notes
3. Reader-facing output — markdown and/or HTML that reads like a book, not a log dump
4. Clear interpretation — distinguish facts from inference, and explain what the team likely learned

In practice, the output should have two lanes:

• Narrative lane — the readable book: phases, pivots, subsystem arcs, pressure loops, lessons
• Evidence lane — the verification layer: commit explorer, raw notes, PR lists, tags, appendices

Non-negotiable rules

• Build the book in passes. First produce durable evidence and notes; then do a second pass to improve readability, chaptering, and UI.
• Keep a narrative lane and an evidence lane. Readers should be able to skim the story or drill into proof.
• Cross-link everything. The final HTML should link back to day notes, phase notes, commit notes, PRs, and raw chapters.
• Persist after every chunk. Never keep important history analysis only in the model context.
• Summarize hierarchically. Commit → chunk/day → phase → book chapter.
• Use stable chunking. Chunks must be deterministic so work can resume later.
• Keep provenance. Every major claim should be traceable to commits/PRs.
• Separate fact from inference. "They changed X" is fact; "they learned Y" is interpretation.
• Prefer breadth first, depth second. First map the whole repo, then deepen important zones.
• Use an orchestration protocol. Large histories should be processed by stable passes and merged through note files, not by trying to remember prior analysis. See references/orchestration-pattern.md.

Workflow

Step 1 — Export the full history

Create a durable workspace, usually under .context/history-book/:

.context/history-book/
├── exports/
│   ├── all-commits.json
│   ├── commits-by-day.json
│   ├── first-parent.json
│   ├── merged-prs.json
│   ├── tags.json
│   └── all-commits.tsv
├── chunks/
│   ├── chunks.json
│   └── chunk-0001.md
├── notes/
│   ├── commits/
│   ├── days/
│   ├── phases/
│   └── pivots/
├── book/
│   ├── index.md
│   ├── timeline.md
│   ├── lessons.md
│   └── appendix-commits.md
└── manifest.json

Use bundled scripts when possible. Paths are relative to the skill directory (e.g. ~/.claude/skills/repo-history-book/ after install, or skills/engineering/repo-history-book/ in this repo):

# 1. Export git history into durable JSON/TSV
python3 "$SKILL_DIR/scripts/export_git_history.py" /path/to/repo --out .context/history-book/exports

# 2. Build stable chunks
python3 "$SKILL_DIR/scripts/chunk_history.py" \
    .context/history-book/exports/commits-by-day.json \
    --out .context/history-book/chunks/chunks.json

# 3. (later, after narrative/days/subsystems/deep_lessons.json exist) render the HTML book
python3 "$SKILL_DIR/scripts/build_book.py" --in .context/history-book --title "My Project"

If the repo has GitHub PR history available (via gh), the export captures it automatically. If gh is absent or unauthenticated, the script silently proceeds with git alone.

Note on the workspace path. .context/history-book/ works out-of-the-box in Conductor (where .context/ is gitignored by default). Outside Conductor, add .context/ to your .gitignore before running, or choose a different output directory.

Step 2 — Create stable chunks

Default chunking strategy:

1. Group by calendar day
2. If a day has more than ~25 commits, split into day-part-a, day-part-b, etc.
3. If the repo is sparse, chunk by 20–30 commits instead

Do not invent ad hoc batches midstream. The chunk IDs must stay stable across resumptions.

See references/chunking-strategy.md.

Step 3 — Write commit notes, not essays

For each commit in a chunk, write a compact structured note using the schema in references/commit-note-schema.md.

Focus on:

• what changed
• which subsystem changed
• whether this was feature / fix / refactor / infra / docs / rollback / brand / release
• why it matters in the broader story
• what pain point or learning it hints at

Keep each commit note short. The goal is durable memory, not prose polish.

Step 4 — Merge commit notes into day notes

After finishing a chunk, immediately write or update:

• notes/days/YYYY-MM-DD.md
• manifest.json

A day note should answer:

• what happened on this day?
• what were the dominant themes?
• what pains were recurring?
• what new capability, refactor, rollback, or belief shift occurred?
• what might the team have learned?

Use references/day-note-schema.md.

Step 5 — Build phase notes

Once day notes exist, group them into 4–8 larger phases. Good boundaries usually come from:

• major PR merges
• renames / rebrands
• subsystem deletions
• deployment model changes
• release-train bursts
• repeated fixes around the same incident class

Phase notes are where the repo starts to become readable as a narrative.

Step 6 — Extract pivots and aha moments

Use the heuristics in references/aha-heuristics.md.

Especially look for:

• add → fix → revert loops
• repeated fixes in the same area within a few days
• deleting an entire subsystem
• adding CI/tests/compliance after instability
• docs/papers/legal pages appearing after technical churn
• renames that clarify positioning
• packaging/install work dominating feature work

These often reveal what the engineers actually learned.

Step 7 — Write the book

The book is assembled from JSON files that live under .context/history-book/book/:

• narrative.json — thesis, phases, pivots, pressure_loops, lessons (card-sized)
• days.json — map of date → day summary (title, summary, themes, kind, key_commits)
• subsystems.json (optional) — array of subsystem arcs
• deep_lessons.json (strongly recommended) — chapter-length deep lessons (see Step 7a)

Once those JSON files exist, run the bundled renderer to produce a self-contained HTML artifact with navigation, search, and a commit explorer:

python3 "$SKILL_DIR/scripts/build_book.py" --in .context/history-book --title "My Project"
# opens: .context/history-book/book/index.html

The script works incrementally — if deep_lessons.json or subsystems.json are missing, those sections are hidden from the output and the nav. Re-run the script any time you update the underlying JSON.

The final book should usually have:

• Overview — one-paragraph read on what this repo became
• How to read this book — explain the narrative lane vs evidence lane
• Timeline — major phases
• Pivots — before/after moments
• Pressure loops — repeated pain patterns like packaging, billing, runtime, or rollback loops
• Lessons at a glance — one-card summaries for skimming
• Deep lessons — full chapters on the load-bearing lessons (see Step 7a)
• Subsystem arcs — how major components changed role over time
• Daily / chunk notes — exhaustive but skimmable
• Appendix — all commits / PRs / tags

See references/book-outline.md.

Step 7a — Write deep lessons

The lesson cards above are easy to skim and easy to forget. The deep lessons chapter is the book's center of gravity — a small number of full chapters on the load-bearing things the team actually learned.

A topic deserves a deep lesson when it has at least two of:

• A dedicated section of the team's own notes (e.g. docs/agents/lessons-learned.md) with 4+ bullets
• A clear pressure loop in the commit record (10+ commits in one area over weeks)
• A hot file (top 5 in top-files.tsv) that keeps being rewritten
• A pivot or architectural bet whose consequences echo through later work

For each deep lesson, write: title (as a claim, not a topic), one quotable one_liner, problem, bulleted what_broke symptoms (grounded in primary sources), what_they_learned (the rule they converged on), transferable (the universal insight), and evidence (PR numbers, file paths, team-doc sections).

Aim for 6–10 deep lessons. Fewer and the chapter is thin; more and the signal dilutes. Use size categories (xl / l / m / s) so the heaviest lessons look heavy.

See references/deep-lesson-schema.md and references/sample-deep-lesson.md.

Two primary sources that will disproportionately feed this chapter:

1. The team's own lessons doc. Most projects that write one use a path like docs/agents/lessons-learned.md, CONTRIBUTING.md, or docs/engineering-notes.md. Grep for file names matching lessons, learnings, postmortem, decisions — if one exists, it is gold and the deep lessons should mostly trace back to its sections.
2. The pressure loops you already identified. Every pressure loop that lasted weeks is almost certainly a deep lesson waiting to be written.

Step 8 — Do a second-pass review

The first build is rarely good enough. After generating the initial book/site, review it specifically for:

• Does it feel like a book, or just a dashboard?
• Is there a clear reading path for humans?
• Are the biggest learning loops obvious?
• Can a reader jump from prose to raw evidence in one click?
• Are the most important subsystem arcs clearly explained?

Use references/output-review-checklist.md. Patch the skill output and rerender.

Interpreting responsibly

Do not overclaim motive. Use careful language:

• Good: "This suggests the team learned…"
• Good: "A plausible interpretation is…"
• Good: "The commit pattern implies…"
• Bad: "The team definitely believed…" unless documented explicitly

When uncertain, name the uncertainty.

Orchestration pattern

For large repos, long-running work, or multi-agent execution, use the merge protocol in references/orchestration-pattern.md. The short version: one pass extracts notes, one pass merges them upward, one pass authors the book/site, and every pass reads from durable files rather than model memory.

Resume strategy

When continuing work later:

1. Read manifest.json
2. Read the latest chunk note and the relevant day note
3. Read only the next unfinished export/chunk files
4. Continue from there

Do not reload the entire book into context unless necessary.

Quality bar

A good result is:

• exhaustive in evidence
• layered in summary
• readable like a book
• honest about uncertainty
• explicit about lessons, pain loops, and pivots

A bad result is:

• one giant undifferentiated summary
• no durable notes
• no stable chunking
• no provenance
• treating every commit as equally important in the prose layer

HTML expectation

When the user asks for HTML, prefer a polished interactive artifact with:

• section navigation
• chapter cards
• lesson cards
• phase timeline
• pivot cards
• pressure-loop cards
• searchable commit explorer
• raw-note backlinks

Use references/html-book-experience.md.

---

AZ Skill Feedback

After the run, if this skill was useful, confusing, broken, or missing context, submit 1-3 safe sentences. Include --agent-harness when known. Include --model-config only if the harness exposes the exact model/config; otherwise leave it as unknown:

python3 ~/.agents/skills/skill-feedback/scripts/skill-feedback.py --skill repo-history-book --rating idea --text "..."

Never include secrets, private data, source code, long prompts, or stack traces.