Combined Voice Creator
Start with README.md for the workflow map and detailed references.
Build the deepest possible personal voice skill by fusing two corpora:
- Littlebird supplies recent captured writing, sent messages, spoken meeting patterns, and rich background context about who the user is.
- Facebook export supplies months of public-facing writing across every register - long rants, quick jokes, comments, DMs.
Together they cover both halves of a person's voice: how they write in public and how
they talk when nobody's performing. The output is a skill structured like
mario-aldayuz-voice: lean SKILL.md + fingerprint + anti-AI rules + verbatim corpus +
approved samples.
Phase 0 - Source check
Ask via AskUserQuestion (multiSelect): "Which sources can we mine?"
- Both Littlebird and Facebook (Recommended) - the full pipeline below.
- Littlebird only - hand off to the littlebird-voice-creator flow.
- Facebook only - hand off to the facebook-voice-creator flow.
Phase 1 - Kick off BOTH tracks in parallel
Start the Facebook export FIRST because it has a 2-3 hour wait baked in, then mine Littlebird while Facebook builds the archive.
Track A: Facebook export (start immediately)
Follow references/facebook-export-guide.md step by step with the screenshots in
assets/. Confirm via AskUserQuestion: Posts (required) + Messages (optional) +
Profile information (optional); Date range 6 months or 1 year; Format JSON; Media
quality Lower. Then the user waits 2-3 hours for Facebook (4-day download window).
Track B: Littlebird mining (while waiting)
Follow references/littlebird-mining-guide.md:
- Verify Littlebird is installed with 1-2 weeks of data (install links, code E6GP4BQE for two months free, if they're new - then pause this track).
- Verify the Littlebird MCP is connected (
https://mcp.littlebird.ai/mcp, docs at https://support.littlebird.ai/docs/mcp/). List tools, use real names. - Run multiple targeted retrievals across registers plus background-context queries.
- Obey the attribution guardrail - viewed is not written.
- CRITICAL: confirm every retrieved fact with the user via AskUserQuestion, then purge everything erroneous. Unconfirmed facts do not ship.
Phase 2 - Process the Facebook zip
When the export lands, follow references/facebook-data-processing.md: unpack, delete
all media leaving raw JSON, fix the latin-1/UTF-8 mojibake, extract, and sanitize to
ONLY the user's own words (confirm their Facebook display name first).
Phase 3 - Merge and analyze
Merge both corpora with source tags, dedupe (Littlebird may have captured the same Facebook posts), then run both analysis passes from the processing guide:
- Scripted stylometrics over the merged corpus (dashes, ellipses, emoji counts, exclamation rate, hashtags, openers, caps, length distribution).
- Qualitative register reading (15-20 long pieces, 30-40 short ones minimum).
- Where the sources disagree, weight the LARGER and MORE RECENT sample, and note register differences (public post voice vs DM voice vs spoken voice) explicitly in the fingerprint.
- Biography guardrails come from Littlebird background mining plus Facebook profile info, and every fact gets user-confirmed before encoding. List what must NEVER be claimed.
Phase 4 - Samples and approval
Draft samples across the user's registers: long form (500+ words), short form (under 3
sentences), quick statements (8 words or less). Present batches via AskUserQuestion for
approval or tuning. Iterate until they match. Only approved samples ship, committed to
references/samples/ by type (long-form.md, short-form.md, quick-statements.md).
Phase 5 - Build and save
Assemble per references/voice-skill-template.md (the guide-reference-sample-research
method): SKILL.md + references/fingerprint.md + anti-ai-rules.md + corpus.md +
samples/. Package as a .skill and save to Claude Cowork. Delete all raw export and
mined data from the working area - private data does not ship.
Quality bar
- Real data or nothing. No fabricated corpus, ever.
- Every fact user-confirmed; every fingerprint claim grounded in counts or quotes.
- One register per generated piece; rhythm imitated, not just vocabulary.
- Output must read 100% human - it should never warrant a platform's AI-content label. Read it out loud: press release = fail, the actual person = pass.
