build-voice-vocabulary

name	build-voice-vocabulary
description	Builds a Voice Vocabulary for an author — a dictionary of their characteristic phrases, turns of speech, fillers, metaphors, and contrarian formulas. Used as a humanization layer on top of the DIP. Collected ONLY from raw sources (transcripts, voice input, personally written Telegram posts), NEVER from the author's LinkedIn posts.
disable-model-invocation	true
argument-hint	["Author"]
allowed-tools	Read, Grep, Glob, Edit, Write

Build Voice Vocabulary

Voice Vocabulary is a layer on top of the DIP (Digital Identity Profile), specifically for humanizing the final draft of a post. The DIP describes tone and topics. Voice Vocabulary provides verbatim phrases from the author that you can use as targeted replacements for AI markers.

When to invoke

• After the author has a DIP at v0.7+ (3+ transcripts in _Sources/).
• On the first humanization attempt — when it becomes clear that DIP quotes aren't enough.
• Every 3 months — refresh with new transcripts.

Pipeline (8 steps)

Step 1: Collect source list

Read everything available about the author:

• Transcripts of calls, podcasts, interviews where the author is a speaker
• Voice input (if audio logs exist)
• Telegram posts personally written by the author (not reposts)
• Internal meetings where the author talks a lot live (NOT a Granola summary — you need the full transcript)
• The author's CV, articles (only if written by them, without AI)

❌ Do NOT use:

• The author's LinkedIn posts (even their own) — that's OUTPUT, feedback loop = voice degradation
• AI-generated content, even if the author approved it
• Other people's posts "in the author's style"
• Transcripts where the author only reacts to others (chat comments during someone else's talk)

Step 2: Spin up the template

Copy Templates/Voice-Vocabulary-Template.md → Knowledge-Base/01-Identity-Profiles/_Voice-Vocabulary/[Author]-Voice-Vocabulary.md.

Fill the frontmatter (author, sources, language coverage, status: v0.1 initial).

Step 3: Extract Sentence Starters

Read each transcript. Find recurring sentence openings (5+ times for high, 3–5 for medium, 1–2 for low).

Record in this format:

| Phrase | Frequency | Example (with source) | When to use |
| "I think I figured out..." | high (10+) | "I think I figured out how to give agents access..." — TG 2026-02-16 | Hedged claim setup |

Start with high-frequency patterns (10+). Then medium. Skip low — it's not a pattern.

Step 4: Idiomatic Phrases

Personal turns of phrase that are recognizable as this author. Often — short 2–4-word expressions with characteristic flavor.

Test: "if I heard this phrase from behind a wall, would I know it's [Author]?"

Step 5: Filler Words / Tics

Filler words in spoken speech. Important: use them carefully in LinkedIn posts — 1 per post, no more. They work as a micro-signal of live speech, but oversaturation becomes irritating.

Step 6: Categories 4–9

Same approach for:

• Characteristic Adjectives & Verbs
• Code-Switching Patterns (if the author is multilingual)
• Numbers & Scale References (how they handle numbers)
• Recurring Metaphors (metaphors with repetition — NOT one-offs)
• Contrarian Move Formulas (formulas like "everyone thinks X, in reality Y")
• Closing Patterns (how they wrap up a thought)

Step 7: Sources section

At the end of the file — a table of all source files with type and date. This lets you:

• Update properly as new transcripts appear
• See which category is under-sampled (e.g., only 1 EN source → English coverage is weak)

Step 8: Status + What's NOT covered

In the footer, set the status (v0.1 / v0.5 / v0.7 / v1.0) and the "Not Covered / TBD" block:

• Categories with insufficient entries
• Hypotheses to verify against future transcripts
• Questions for the author

Validity checklist

After completing the first pass, walk through the checklist:

• Minimum 3–5 entries in each main category (1–4, 9)?
• Every entry has a verbatim quote + source ref?
• Frequency is correctly assigned (not "high" for everything)?
• At least 5 distinct files covered in Sources?
• No entries from the author's LinkedIn posts (raw sources only)?
• Frontmatter fully filled?
• Status is accurate (if you only have 3 sources → v0.5, not v1.0)?

If it fails on ≥2 items — that's v0.5 (working), not v1.0.

Anti-patterns

• ❌ Rebuilding the DIP inside the Voice Vocabulary — overlapping work.
• ❌ Taking "unique" words from a single transcript as a pattern — that's noise, not signal.
• ❌ Including phrases the author used in their first-ever public talk — they may sound unnatural there.
• ❌ Auto-generating from posts — that produces a feedback loop and loss of voice.

Output

A file at Knowledge-Base/01-Identity-Profiles/_Voice-Vocabulary/[Author]-Voice-Vocabulary.md, ready for use by /detect-ai and /check-generic validators as a reference for humanization replacements.