---
name: facebook-voice-creator
license: AGPL-3.0-or-later
description: "Build a personal writing-voice skill from a Facebook export. Use when the user supplies their own export. Read README.md for the guide map."
metadata:
  hive-tier: standalone
  version: "1.0.0"
  author: "Mario Aldayuz / Littlebird"
---

# Facebook Voice Creator

Start with [README.md](README.md) for the workflow map and detailed references.

Build a user's personal voice skill from their real Facebook writing. The output is a
skill like `mario-aldayuz-voice`: a lean SKILL.md plus fingerprint, anti-AI rules,
verbatim corpus, and approved samples - all derived from what the user actually wrote,
never from guesses.

Why Facebook data: a data export contains months of the user's unfiltered writing across
every register they have - long rants, quick jokes, DMs, comments. That beats any
interview or questionnaire. The whole method is: real words first, analysis second,
imitation last.

## Process overview

Run these phases in order. Use the AskUserQuestion tool at every decision point - the
reference guides specify each question and its recommended answer.

### Phase 1 - Export (user-driven, ~5 minutes of clicking)

Read `references/facebook-export-guide.md` and walk the user through the Facebook
export flow step by step, showing them the screenshots in `assets/` as you go.

Critical settings to confirm with AskUserQuestion (recommended answers first):
1. Customize information: **Posts (required)**, Messages (optional), Profile
   information (optional). Nothing else.
2. Date range: **6 months** or 1 year.
3. Format: **JSON** (not the HTML default - this matters).
4. Media quality: **Lower**.

### Phase 2 - Wait (async, 2-3 hours)

Facebook builds the export in roughly 2-3 hours and the user has only 4 days to
download it. Do not block - tell the user to return with the zip. If there is other
work in this conversation, proceed with it.

### Phase 3 - Process the upload

When the user uploads the .zip, read `references/facebook-data-processing.md` and
follow it exactly: unpack, delete all media leaving only raw JSON, fix the latin-1/UTF-8
mojibake, extract text, and SANITIZE - keep only content written by the user themselves
(confirm their exact Facebook display name first). Nothing from other people's comments
or messages may enter the corpus.

### Phase 4 - Analyze

Run both analysis passes from the processing guide: scripted stylometrics (dash usage,
emoji counts, exclamation rate, hashtags, openers, caps, lengths) and qualitative
register reading of at least 15-20 long posts and 30-40 short ones. Confirm biography
facts with the user before encoding them as guardrails - never encode unverified claims
(service history, credentials, life events).

### Phase 5 - Samples and approval

Draft sample posts in the user's voice: long form (500+ words), short form (under 3
sentences), and quick statements (8 words or less). Present them via AskUserQuestion
for approval or tuning. Iterate until the user signs off. Only approved samples ship.

### Phase 6 - Build and save the skill

Assemble the final skill using `references/voice-skill-template.md`: SKILL.md plus
references/fingerprint.md, anti-ai-rules.md, corpus.md, and samples/ split by type
(long-form.md, short-form.md, quick-statements.md). Save the skill to Claude Cowork
(package as .skill and present it with a Save skill button). Delete the raw export
data - it is private and does not belong in the shipped skill.

## Quality bar

- Every fingerprint claim is grounded in counts or quotes from the real corpus.
- Every biography fact is user-confirmed. The guardrails list what must NEVER be
  claimed.
- The calibration test lives in the output skill: read it out loud - if it sounds like
  a press release it failed, if it sounds like the user it passed.
- The finished skill must produce writing that reads 100% human and never warrants a
  platform's AI-content label.