Voice Builder

Use to build a reusable voice guide (voice.md) by analyzing a person's or brand's ACTUAL writing samples — so every other social skill can write in their exa...

installs

stars

karma

SkillRank score ↗

8.3/ 10

evaluated by implexa, claude-haiku-4-5 · 2026-07-22

voice-builder extracts reproducible writing patterns from real samples to produce a voice.md that downstream skills can write from. it works backwards from evidence (quotes, syntax patterns, negative space) rather than adjectives, then validates the result before shipping.

structure

9.0

trigger phrases

9.0

procedure

9.0

edge cases

8.0

documentation

8.0

view original SKILL.md from clawhubclick to expand

---
name: voice-builder
description: >-
Use to build a reusable voice guide (voice.md) by analyzing a person's or brand's ACTUAL
writing samples — so every other social skill can write in their exact voice instead of
generic AI prose. Run this when the user says "build my voice," "sound like me," "capture my
tone," "analyze my writing," "train on my posts," "ghostwrite in my voice," or when content
keeps coming out off-voice or generic. Requires real writing samples; if there are none,
use brand-profile's voice interview instead. Works for anyone: founders, creators, executives
who are ghostwritten, and brands. Read brand-profile first; this skill produces the deeper
voice.md that complements it. This skill DEFINES the voice; to apply it per piece — tone
shifts, style edits, de-AI-ing a draft — use writing-style-and-tone.
metadata:
version: 1.0.0
license: MIT
---

# Voice Builder

This skill turns real writing into a **reproducible voice** — a `voice.md` that any
downstream skill (or human) can write from and produce posts indistinguishable from the
source.

The core principle: **voice is extracted from evidence, not described in adjectives.**
"Punchy and authentic" is unwritable — it tells a writer nothing. "Opens with a one-line
provocation, writes in short sentences with one long one for contrast, uses em-dashes, never
uses exclamation marks, ends on a question" is *reproducible*. Your job is to convert someone's
writing into rules that specific.

## When to use this

- **After `brand-profile`**, to go deep on voice when the user has writing samples.
- When content keeps coming out generic, stiff, or "not me."
- For **ghostwriting at scale** — capture the voice once, write in it forever.

**When NOT to use this:** if there are no usable samples, don't invent a voice. Route to
`brand-profile`'s voice interview (dimensions + lexicon + the "we sound ___, never ___" line)
and come back when samples exist. A fabricated voice is worse than an honest interview.

## Step 0 — Read `brand-profile` first

Load `brand-profile.md` for identity, audience, and guardrails. `voice-builder` produces the
`voice.md` that sits alongside it. If a `voice.md` already exists and is current, load it,
summarize it back, and skip to whatever the user actually needs.

## Step 1 — Gather the right samples

Quality beats quantity. The best samples are **unmistakably the person** — pieces they're
proud of, in their natural register, that they actually wrote (not committee-edited, not
ghostwritten by someone else, not AI-generated). Aim for **5–10 substantial samples**; 3 is
the floor. **Spoken transcripts are gold** — they capture natural rhythm before self-editing
flattens it. See `references/samples-guide.md` for what to collect and what to exclude.

If samples are thin or inconsistent, say so plainly and proceed at lower confidence rather
than overstating. Honesty about confidence is part of the deliverable.

## Step 2 — Analyze across the six layers (with evidence)

Do not free-associate about "tone." Work the framework in
`references/analysis-framework.md`, layer by layer:

1. **Lexicon** — word choice, signature phrases, contractions, register, banned-by-habit words.
2. **Syntax** — sentence length and variance, openers, fragments, active/passive, rhythm.
3. **Rhetoric** — questions, direct address, repetition, analogy, contrast, humor, the "turn."
4. **Structure** — how they open, build, and close; paragraphing; formatting.
5. **Stance** — distance from reader, authority posture, confidence, emotional register.
6. **Negative space** — what they conspicuously *never* do. Often the most defining layer.

For every pattern you name, **cite the evidence** — quote the sample. A voice guide built on
assertions drifts; one built on quotes holds.

## Step 3 — Distill the voice fingerprint

From the analysis, extract the **5–7 most distinctive, reproducible signatures** — the few
things that, done right, make the writing recognizable as theirs. This is the heart of
`voice.md`. Pair it with the **negative space** (the hard "never"s). If you nailed only the
fingerprint, a reader should already believe it's them.

## Step 4 — Write `voice.md`

Use `references/voice-template.md`. Make it **operational**, not a personality essay: rules a
writer can follow, a tone map for how the voice flexes by context, and 3–5 **do/don't
rewrites** (a generic sentence → the same thing in this voice). Keep real sample snippets in
as calibration anchors.

## Step 5 — Validate with the voice test (do not skip)

A `voice.md` that hasn't been tested is a hypothesis. Run the loop in
`references/validation.md`:

1. Generate a fresh post on a *new* topic using only `voice.md`.
2. Place it beside a real sample. Could the user tell which is which?
3. Score it against the fingerprint signatures — which did it hit, which did it miss?
4. Fix the misses, refine `voice.md`, and repeat once or twice.

Show the user the test post and the result. The voice isn't done until it passes.

## What "great" looks like (self-check before finishing)

- A writer who has never met this person could write an on-voice post from `voice.md` alone.
- **Every claim cites a sample** — no unsupported adjectives.
- The fingerprint is specific enough to be *falsifiable* (you could point to a post and say
"that's not them, because…").
- Negative space is captured, not just positive traits.
- It **passed the voice test**, and you showed the proof.

If any of these is missing, you're not done.

## Edge cases — handle explicitly

- **Thin samples (1–2 short pieces):** capture what's observable, mark the guide
**low-confidence**, name which layers you couldn't determine, and supplement with the
`brand-profile` interview. Don't bluff certainty.
- **Inconsistent samples (voice drift):** don't average them into mush. Flag the inconsistency
and ask which samples represent the *target* voice; anchor on those.
- **Aspirational voice** (they want to sound different than they currently do): separate
**current** from **target**. Collect samples of who they want to sound like, label them
aspirational, and blend transparently so the user knows what's being borrowed.
- **Multiple voices** (company vs founder; several creators): one `voice.md` per voice, named.
**Never blend voices** — it produces a person who exists nowhere.
- **Heavily edited / ghostwritten / AI-generated samples:** exclude them; they aren't the
person's voice. Say why.
- **Spoken vs written:** transcripts reveal natural rhythm but need cleanup rules (filler,
false starts). Capture the rhythm, note the cleanup.
- **Multilingual:** analyze per language; voice rarely maps 1:1 across languages.

## Related skills

- `brand-profile` — read first; provides identity, audience, guardrails.
- `writing-style-and-tone` — applies the `voice.md` per piece and per moment (tone map, edit
passes, AI-tell sweep). This skill writes the constitution; that one governs by it.
- Every content skill (`caption-writer`, `reels-script`, `thread-writer`, `linkedin-post-writer`, …)
reads the resulting `voice.md`.
- `audience-research`, `content-pillars` — complementary foundation skills.

## References

- `references/samples-guide.md` — how to source and vet the right writing samples.
- `references/analysis-framework.md` — the six-layer framework for deconstructing voice.
- `references/voice-template.md` — the exact `voice.md` structure to produce.
- `references/validation.md` — the voice test: generate, compare, score, refine.
- `references/examples.md` — two contrasting worked examples, end to end.

don't have the plugin yet? install it then click "run inline in claude" again.

added explicit inputs (brand-profile.md, references, optional aspirational context), extracted hidden decision logic into nine if-else branches (missing brand-profile, existing voice.md, thin samples, inconsistency, aspirational voice, multiple voices, edited samples, transcripts, multilingual), clarified procedure as six numbered steps with input/output, documented edge cases as decision points, specified output contract (voice.md location, format, citation requirement) and outcome signal (testable, usable, traceable).

Voice Builder

intent

build a reproducible voice guide (voice.md) by deconstructing real writing samples into explicit, actionable rules. run this when the user says "build my voice," "sound like me," "capture my tone," "analyze my writing," "train on my posts," or when content keeps coming out generic and off-brand. works for founders, creators, executives who are ghostwritten, and brands. the output lives alongside brand-profile.md and serves as the source of truth for every downstream content skill. skip this if you have no writing samples; route to brand-profile's voice interview instead. a fabricated voice is worse than an honest interview.

inputs

brand-profile.md (required): load existing brand identity, audience, and guardrails. if voice.md already exists and is current, load it and summarize back to the user before proceeding. if neither exists, run brand-profile first.
writing samples (required): 5-10 substantial pieces the user actually wrote (not committee-edited, not ghostwritten by someone else, not AI-generated). spoken transcripts are gold; they capture rhythm before self-editing flattens it. minimum floor is 3 samples. see references/samples-guide.md for sourcing and vetting.
references/analysis-framework.md: the six-layer deconstruction template (lexicon, syntax, rhetoric, structure, stance, negative space).
references/voice-template.md: the voice.md output structure.
references/validation.md: the voice test protocol (generate, compare, score, refine).
references/examples.md: two worked examples for calibration.
optional context: if the user has an aspirational voice target, collect samples of who they want to sound like; label them aspirational and blend transparently.

procedure

Step 1: Load brand-profile and check for existing voice.md

Input: user's request to build voice.
Load brand-profile.md. If voice.md already exists and is current, summarize it back and ask what the user actually needs (refinement, validation, application).
If brand-profile.md doesn't exist, tell the user to run brand-profile first. do not proceed without identity, audience, and guardrails.
Output: confirmation of brand context; proceed flag or redirect.

Step 2: Gather and vet writing samples

Input: user's writing samples (links, pasted text, document uploads, transcripts).
Apply vetting criteria from references/samples-guide.md: unmistakably the person, pieces they're proud of, natural register, actually written by them (not AI, not ghostwritten). exclude heavily edited, committee-output, and AI-generated samples; say why.
If samples are thin (1-2 short pieces) or inconsistent (voice drift across samples), flag confidence level now. don't bluff certainty.
If multiple voices exist (company vs founder, multiple creators), establish one voice.md per voice, named separately. never blend voices.
Output: vetted sample set with quality assessment and confidence note.

Step 3: Analyze across six layers with evidence

Input: vetted samples, references/analysis-framework.md.
Work layer by layer. for each pattern, cite evidence (quote the sample). do not free-associate.
- Lexicon: word choice, signature phrases, contractions, register, banned-by-habit words.
- Syntax: sentence length and variance, openers, fragments, active/passive, rhythm.
- Rhetoric: questions, direct address, repetition, analogy, contrast, humor, the turn.
- Structure: how they open, build, and close; paragraphing; formatting choices.
- Stance: distance from reader, authority posture, confidence, emotional register.
- Negative space: what they conspicuously never do (often the most defining layer).
Output: analysis memo with quoted evidence for each layer.

Step 4: Distill the voice fingerprint

Input: analysis memo, confidence level from step 2.
Extract 5-7 most distinctive, reproducible signatures. pair with negative space (hard "never"s). if you nailed only the fingerprint, a reader should already believe it's them.
Output: fingerprint list with supporting quotes.

Step 5: Write voice.md

Input: fingerprint, negative space, references/voice-template.md.
produce operational rules, not personality essay. include tone map (how voice flexes by context), 3-5 do/don't rewrites (generic sentence to this voice), and real sample snippets as calibration anchors.
Output: draft voice.md.

Step 6: Validate with the voice test (do not skip)

Input: draft voice.md, references/validation.md.
run the loop:
- Generate a fresh post on a new topic using only voice.md (no user input beyond topic).
- place it beside a real sample. could the user tell which is which?
- score against fingerprint signatures: which did it hit, which did it miss?
- fix misses, refine voice.md, repeat once or twice.
show user the test post and validation result.
Output: tested and refined voice.md, validation proof (test post + comparison).

decision points

if brand-profile.md does not exist: stop. tell user to run brand-profile first. voice requires identity, audience, and guardrails before it can be built.

if voice.md already exists and is current: load it, summarize back to user, skip to step 6 (validation/refinement) or ask what they actually need.

if samples are fewer than 3 or very short: proceed at low confidence. mark the guide low-confidence, name which layers you couldn't determine, suggest supplementing with brand-profile interview. do not fabricate certainty.

if samples are inconsistent (voice drift across time): don't average them into mush. flag the inconsistency. ask which samples represent the target voice (current or aspirational). anchor analysis on target samples.

if user wants aspirational voice (sound different than currently does): separate current from target. collect samples of who they want to sound like, label aspirational, blend transparently in voice.md so user knows what's borrowed. output both current and target, labeled.

if multiple distinct voices exist (company vs founder, multiple creators): one voice.md per voice. name each clearly. never blend voices into a hybrid that exists nowhere.

if samples are heavily edited, ghostwritten, or AI-generated: exclude. tell user why (not the person's voice). ask for replacements.

if samples are spoken transcripts: transcripts reveal natural rhythm but need cleanup rules (filler, false starts, verbal tics). capture the rhythm, note cleanup rules so downstream skills know when to strip them.

if samples are in multiple languages: analyze per language. voice rarely maps 1:1 across languages. produce one voice.md per language.

if voice test fails (generated post doesn't sound like them): refine voice.md and test again. repeat once or twice. if it still fails after two iterations, flag that this voice may not be learnable from current samples; suggest collecting more samples or running brand-profile interview to supplement.

output contract

voice.md file (required):

location: same directory as brand-profile.md, named voice.md.
format: markdown with sections for fingerprint, negative space, tone map, do/don't rewrites, real sample snippets, confidence note if applicable.
structure: follows references/voice-template.md exactly.
every claim cites a sample (no unsupported adjectives).
fingerprint is specific enough to be falsifiable (you could point to a post and say "that's not them, because…").
includes tested, validated version with proof of voice test (test post + comparison).

validation memo (required):

test post generated on a new topic using only voice.md.
side-by-side comparison with real sample.
scored against fingerprint signatures (hit/miss per signature).
confidence assessment and any refinement notes.

edge case documentation (if applicable):

low-confidence note if samples are thin or inconsistent.
aspirational vs current labels if dual voice.
cleanup rules if transcripts are sourced.
language note if multilingual.
exclusion notes for any samples rejected (why they weren't used).

outcome signal

voice.md is written and tested: the user has a reproducible voice guide ready for downstream skills.
voice test passes: a generated post on a new topic is indistinguishable from a real sample, or user confirms it captures their voice.
every other content skill can now read voice.md and write on-voice posts without user intervention per piece (writing-style-and-tone applies it; caption-writer, reels-script, thread-writer, linkedin-post-writer, etc. all read it).
user confirms the fingerprint: when read back, user says "yes, that's me" or points to specific refinements. refinements are incorporated and re-tested.
no unsupported claims: every assertion in voice.md is backed by a quoted sample. user can trace any rule back to evidence.

Voice Builder

related skills

Voice Builder

intent

inputs

procedure

decision points

output contract

outcome signal