Voice File Builder

---
name: voice-file-builder
description: Extract a person's writing voice from their real published work into an enforceable voice-rules file an agent loads before writing anything. Use when asked to make the AI write like me, build a voice or style guide for an agent, fix content that sounds generic or like AI, capture a founder's tone, keep ghost-written content consistent, or someone says 'this doesn't sound like us at all'. For where the file lives, see agent-folder-scaffold. For the driver that points to it, see claude-md-writer.
metadata:
  version: 1.0.0
---

# Voice file builder

The course's brief is one sentence: *"Be specific enough that another writer
could pass as me."* A voice file is not adjectives ("professional yet
approachable") — it is **rules extracted from real writing**, loaded by the
writing skill first, every time.

The input rule that decides everything: the file is built from **pieces the
person actually wrote themselves** — not from what they wish they sounded like.

## Before you start

| Input | Why |
| --- | --- |
| 3+ pieces the person genuinely wrote | The corpus prompt starts "Here are three posts I wrote myself" — ghost-written pieces extract the ghost's voice |
| Which content types the voice covers | A LinkedIn voice and a whitepaper voice may differ; say which this file governs |
| Anything they refuse to sound like | The banned list is half the file's power |

## Step 1 — Extract the four sections

From the real writing, in the course's structure:

1. **Banned phrases** — words and constructions this person never uses. Include
   the AI tells they care about (the course bans em dashes outright: "the most
   recognisable AI tell in written English").
2. **Sentence rhythm, with real examples** — short/long patterns, fragment use,
   paragraph length, *quoted from their pieces*. Rules without examples drift;
   examples are what another writer imitates.
3. **Personal-proof patterns** — how they cite their own experience and numbers:
   "we tested this with 40 clients", "my CPL dropped 23%". The shape, plus which
   real numbers they may draw on.
4. **The imperfect-truths rule** — what keeps the copy human: admitted failures,
   hedged claims where honesty demands it, the things they say that polish would
   remove.

## Step 2 — Make every rule checkable

| Adjective (useless) | Rule (enforceable) |
| --- | --- |
| "Conversational tone" | Sentences average under 15 words; fragments allowed; one-line paragraphs used for emphasis |
| "Confident but humble" | States numbers plainly; admits one limitation per piece; never "revolutionary/game-changing" |
| "Professional" | No emoji in body copy; no exclamation marks in claims |

An editor should be able to mark any paragraph pass/fail against the file.

## Step 3 — Wire it into the agent

- File sits next to the writing skill (`voice-rules.md` beside the writer's
  SKILL.md) and the skill **loads it first, every time** — the course's
  blog-content-writer does exactly this.
- CLAUDE.md points to it by name; the rules themselves live here, not in the
  driver.
- Content types it does not govern are named at the top, so nobody applies the
  LinkedIn voice to the legal page.

## Step 4 — Verify by imitation

The acceptance test is the course's own: hand the file (without the samples) to
another writer — or a fresh session — and have them write 150 words on a topic
the person never covered. If a colleague cannot tell it from the real thing, the
file works. If they can, the tell they name becomes the next rule.

## Output

```
# voice-rules.md: <person> · governs: <content types>

## Banned  (<n> entries, incl. em dashes if adopted)
## Rhythm  (rules + quoted examples from their real pieces)
## Proof patterns  (shapes + the real numbers available)
## Imperfect truths  (what stays human)

Built from: <the 3+ self-written pieces, listed>
Imitation test: <pass / the tell found → new rule added>
Loaded by: <which skill loads it first>
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Output still sounds generic | File is adjectives, not rules | Step 2: every rule checkable, with quoted examples |
| Voice drifts across a long piece | File not loaded first by the writing skill | Wire the load order into the skill itself |
| Sounds like the person's ghostwriter | Extracted from ghost-written pieces | Rebuild from genuinely self-written material only |
| Reads human but claims are wrong | Proof patterns copied without the real numbers | The file lists which numbers exist; claims trace there |
| Every piece admits a failure identically | Imperfect-truths applied as a template | It is a licence, not a quota — vary or omit |
| Two team voices blur together | One file governing everyone | One person, one file; a house voice is its own file |
| File approved, never used | Lives in a doc, not next to the skill | voice-rules.md beside the writer skill, loaded first |

Never build the file from an interview about how they'd like to sound.
Aspiration produces adjectives; only real writing produces rules — extract, then
let them veto.

## Rules

- **Real self-written pieces only**, because the file extracts what is actually
  there, and a ghost's voice is the ghost's.
- **Rules with quoted examples, never adjectives**, because "conversational"
  cannot be imitated and a quoted rhythm can.
- **The banned list is half the file**, because voice is defined as much by what
  never appears as by what does.
- **Loaded first by the writing skill, every time**, because a voice file the
  writer skill doesn't load is a wish.
- **Verify by imitation, not approval**, because the owner approving their own
  flattering file proves nothing a colleague's blind read wouldn't disprove.

## Related skills

- **claude-md-writer** — the driver points here; two voice lines max live there.
- **skill-writer** — the writing skill that loads this file first.
- **ai-content-qc** — the human gate checks against this file.
- **anti-ai-copy** (on the site) — the de-AI-ing rules this file personalises.