book-bootstrap

book-bootstrap — auto-scaffold a Hebrew book project

When to invoke

• Beginning of every /start <action> invocation.
• Direct user request to set up a project from an existing manuscript.

Inputs

• Current working directory (the user's project root).
• Optional pre-existing book.yaml, AUTHOR_VOICE.md, .book-producer/.

Process

Step 1 — Detect manuscript files

Scan in this order; first hit wins:

1. chapters/*.md (multi-chapter project — already organised).
2. Files matching ch*.md, chapter-*.md, [0-9][0-9]-*.md, [0-9]-*.md in cwd (multi-chapter, no folder).
3. Single manuscript.md / book.md / largest *.md over 5 KB.
4. *.docx and *.pdf files. For DOCX/PDF, extract text into a temp chapters/ directory and treat each as a chapter.
5. If nothing found and the user said "ספר חדש" / init, scaffold a blank project with one stub chapter.
6. If nothing found and the user wanted a terminal action — abort with one Hebrew error: "לא מצאתי כתב יד. שמור manuscript.md בתיקייה ונסה שוב."

Step 2 — Guess genre from content

Sample the first 3,000 words of the longest detected file. Score:

Signal	Pattern	Genre
Hazal-citation density > 1 per 500 words	בבלי, ירושלמי, מדרש, רמב"ם, שו"ע, (א-ת+)\s״-style	religious
First-person ratio > 0.10 of all words	אני / אנחנו / לי / לנו / שלי	autobiography
Argumentation markers > 5 instances per 1000 words	ברם, אולם, שכן, הואיל ו, מכאן ש, כפי ש	philosophy
None of the above prominent	—	popular-science
Two or more strongly	—	mixed

Step 3 — Auto-create book.yaml

If book.yaml already exists, skip this step. Otherwise write:

# Auto-generated by book-bootstrap on <ISO-date>. Edit any field freely.
title: "<derived from H1 of longest file, or directory name>"
author: "<git config user.name, fallback to empty>"
genre: <guessed genre>
target_words: <sum of detected word counts × 1.1, rounded to nearest 1000>
citation_style: chicago-author-date
niqqud: <true if religious + niqqud detected in sampled text, else false>
deadline: ""
created: <YYYY-MM-DD>
chapters:
  # one entry per detected file, with stage: drafted
  - id: ch01
    file: <relative path>
    stage: drafted
  - id: ch02
    file: ...
    stage: drafted

Chapter IDs follow ch01, ch02, … unless a filename already encodes an order (03-foo.md → id: ch03, preserve it).

Step 4 — Auto-create .book-producer/

If missing:

.book-producer/
  state.json            # { "chapters": [<one per book.yaml chapter, stage: drafted>], "last_update": "<ISO>" }
  memory.md             # header only
  snapshots/            # empty

Update .gitignore to add .book-producer/snapshots/, .book-producer/memory.md, .book-producer/state.json, .ctx/, chapters/*.draft*.md, chapters/*.decisions.md if not already present.

Step 5 — AUTHOR_VOICE.md

If missing:

• Run the express-voice sub-skill (3 questions). It writes AUTHOR_VOICE.md and a sparse .book-producer/profile.json.
• Add a final stub line to AUTHOR_VOICE.md: "להעמקת הקול הרץ /init-voice בכל זמן."

If present:

• Leave it alone. Read it for downstream consumers.

Step 6 — Cache the CandleKeep references

Normally the plugin's SessionStart hook has already cached writers-guide.md, agent-team-guide.md, and hebrew-linguistic-reference.md under .ctx/. If any of those files is missing, fall back to:

bash ${CLAUDE_PLUGIN_ROOT}/scripts/load-candlekeep-guide.sh

The loader is idempotent and handles its own freshness check.

Step 7 — Confirmation (ONE Hebrew question)

After all the auto-creates, summarise in a single sentence and ask one question:

"זיהיתי {N} פרקים, סוגה: {genre}, ניקוד: {on/off}, יעד: {target_words} מילה. נכון? (כן / שנה / ביטול)"

• כן → return success to caller.
• שנה → ask one follow-up: "מה לשנות? (סוגה / ניקוד / יעד / שם)" and let the user override exactly one field. Re-confirm.
• ביטול → roll back any files created in this run (track them in a transactional list during step 3–4) and abort.

Output (returned to the caller)

{
  "status": "ok | aborted | error",
  "scaffolded": ["book.yaml", ".book-producer/state.json", "AUTHOR_VOICE.md"],
  "preExisting": ["book.yaml"],
  "manuscriptFiles": ["chapters/ch01.md", ...],
  "genre": "philosophy",
  "niqqud": false,
  "summary": "<one Hebrew sentence>"
}

Hard rules

• Idempotent. Re-running on a fully scaffolded project must succeed without changing anything (other than possibly refreshing the CandleKeep cache).
• One confirmation question. Never more than one. Bootstrap is for users who don't want to think.
• Never overwrite an existing book.yaml or AUTHOR_VOICE.md. If both already exist, skip steps 3 and 5 silently and proceed.
• The git config name fallback — if git config user.name returns empty, set author: "" and surface that in the confirmation summary so the user knows to fill it.
• DOCX/PDF extraction is best-effort. If extraction yields garbage (Hebrew ratio < 0.5), abort with a Hebrew message asking the user to provide a .md instead.
• No new dependencies. Reuse pdfplumber/python-docx already used by extract-voice-fingerprint.py.

What this skill does NOT do

• It does not run /proof, /edit, etc. — that's the caller's job.
• It does not modify a manuscript's prose. Read-only on the user's content; write-only on metadata files.
• It does not run a full /init-voice. Only express-voice (3 questions). The author can run /init-voice later for the heavy fingerprint.