express-voice

express-voice — minimum-viable voice fingerprint

Why this exists

The full /init-voice flow asks 10 questions and runs a computational fingerprint over past-books/. That's the right depth for an author who's settling in. But for a user who just typed "תוכל להגיה את הספר שלי?" and wants results, 10 questions is a wall.

This skill collects the three highest-leverage voice signals in under a minute. The proofreader and linguistic-editor produce useful output even from this sparse input — what they really can't do is operate with AUTHOR_VOICE.md missing, which would disable voice-preserver entirely.

When to invoke

• book-bootstrap step 5, when AUTHOR_VOICE.md is missing.
• The user explicitly says "רוצה לדלג, רק קצר" during a voice flow.

The three questions

Ask one at a time, in Hebrew, conversational tone. Wait for each answer before the next.

Q1 — Persona

"במשפט אחד, מי המספר של הספר הזה ולמי הוא מדבר?"

→ Stored verbatim under ## Persona.

Q2 — Register

"באיזה משלב? אקדמי, ספרותי, יומיומי, או מעורב?"

→ Stored under ## Register. Map the answer to the five-tier taxonomy from hebrew-linguistic-reference.md chapter hebrew-author-register (high / academic / journalistic / everyday / colloquial). Multi-word answers → mixed with the dominant first.

Q3 — One banned phrase

"ביטוי אחד שאתה לא תכתוב לעולם, שמעצבן אותך לראות בטיוטה?"

→ Verbatim under ## Banned phrases. This single entry gates the linguistic-editor immediately; many authors find that one explicit ban catches 30% of AI-flat fixes on its own.

Output — AUTHOR_VOICE.md

# AUTHOR_VOICE.md

> _שורה זו נוצרה בקצר עם express-voice. הרץ /init-voice בכל זמן להעמקה מלאה (10 שאלות + fingerprint חישובי)._

## Persona
<verbatim Q1 answer>

## Register
- Default: <mapped Q2 — high | academic | journalistic | everyday | colloquial | mixed>
- Switches: _TBD — להשלים עם /init-voice_

## Preferred phrases
_TBD — להשלים עם /init-voice_

## Banned phrases
- <verbatim Q3 answer>

## Sentence rhythm
_TBD — להשלים עם /init-voice. ברירת מחדל: מעורב._

## First person
_TBD — להשלים עם /init-voice._

## Reference paragraphs
_TBD — להשלים עם /init-voice._

## What never to touch
_TBD — להשלים עם /init-voice._

Output — .book-producer/profile.json

If a manuscript exists in the current project (book-bootstrap detected at least one file), also run the extractor on a sample to seed computational fields:

python3 $CLAUDE_PLUGIN_ROOT/scripts/extract-voice-fingerprint.py \
  --input <one of the manuscript files> \
  --baseline .ctx/hebrew-linguistic-reference.md \
  --output .book-producer/profile.json

This gives downstream agents real burstiness, sentence-length, and TTR numbers — even when the author hasn't done the heavy interview. If extraction fails or no manuscript exists, write a stub:

{
  "version": "0.4.0",
  "createdAt": "<ISO>",
  "path": "express",
  "styleFingerprint": null,
  "bannedPhrases": ["<Q3 verbatim>"],
  "preferredPhrases": [],
  "register": "<Q2 mapped>",
  "qualitativeAnalysis": {
    "persona": "<Q1 verbatim>"
  },
  "_note": "express-voice run; upgrade with /init-voice for full fingerprint"
}

Hard rules

• Three questions, no more. If the user volunteers extra info ("גם, אני שונא להשתמש ב X"), capture it as additional banned phrases — but never ask a fourth scripted question.
• Verbatim storage. Don't paraphrase the author's own words. Their phrasing IS the voice.
• Hebrew throughout. Both questions and stored answers.
• Don't overwrite. If AUTHOR_VOICE.md already exists when this skill runs, abort and tell the caller — book-bootstrap should have skipped this skill.

Reporting back

One-line return to the caller:

"הקול נשמר. {persona-1-word}, משלב {register}, ביטוי אסור אחד."