tone-matcher

---
name: tone-matcher
description: Rewrites a draft so it reads like you wrote it, from 3 to 5 samples of your own writing plus the public list of phrases that give AI text away. Use before anything goes out in your name.
---

# Make anything sound like you wrote it

Paste a draft, get it back in your voice, with the words that give AI away removed.

## Claude does

1. **Asks for 3 to 5 samples** you really sent to a real person. Under 3, Claude stops and asks again: Anthropic's rule is 3 to 5 examples, relevant, diverse, structured. One short, one long, one written in a hurry.
2. **Asks who reads it and whose name signs it.** Claude never assumes the sender.
3. **Builds a voice card of countable features, not adjectives:** mean and longest sentence length, contractions, "I" against "we", opener and sign-off, questions per 100 words, bullets or not, five words you use, five you never use.
4. **Loads the tell list** from Wikipedia's Signs of AI writing, the catalogue real editors use to spot AI text: delve, tapestry, testament, underscore, pivotal, boasts, nestled, vibrant, showcasing, meticulous, landscape; "serves as", "stands as", "represents" in place of "is"; "not just X, but Y"; trailing participle clauses; heavy dashes and bold; the "Challenges and Future Directions" wrap-up.
5. **Gives the reason with every rule.** "Never use ellipses" gets ignored; "ellipses read as hesitation, and you do not hesitate" gets followed.
6. **Writes the instruction as actions, not bans:** Rewrite, Shorten, Replace, Keep. Bans are for the tell list only.
7. **Sets sampling:** temperature 0.2, top-P 0.95, top-K 30. Matching a voice is not invention, so 0.9 is wrong.
8. **Fact pass on its own output:** any name, number, date or quote not in your draft is deleted, not guessed.

## Then Codex does

Codex applies the voice card to a folder of drafts, then scans for tells and reports hits and line numbers per file. Bulk matching and rewriting, no judgement needed.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills in first: {VOICE_CARD}, {DRAFT_FOLDER}, {BANNED_TERMS}, {SIGNER_NAME}.

## Claude checks

1. **Control first.** Codex must scan a seeded decoy holding three known tells and report all three. Zero hits proves nothing if the scanner never ran.
2. **Tell count is zero,** and Claude reads two files itself, not just the count.
3. **No new facts.** Every number, name, date and quote in the rewrite appears in the source. One invention rejects the batch.
4. **Voice card match:** mean sentence length within 15 per cent of the card, contractions and person as set.
5. **Signature line present** on every file.
6. **Reject and rerun:** a failed check returns to Codex once, file and line quoted. Fails twice, Claude rewrites that file itself.

## Rules

- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- Three samples minimum, or no voice card.
- Removing a tell never changes a meaning. If the sentence needs the word, keep it.
- Bold and dashes are style. Match the samples, not the model.

## If Codex is not installed

Claude does the rewrite and the scan itself, file by file, and says so.

## Built from

- Wikipedia: Signs of AI writing, https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing, the list editors police AI text with.
- anthropics/skills, https://github.com/anthropics/skills, 174,340 stars from api.github.com.
- Anthropic prompting best practices, https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices, the examples rule.
- Google Prompt Engineering whitepaper, Boonstra, https://www.kaggle.com/whitepaper-prompt-engineering, the sampling numbers.

# Make anything sound like you wrote it - Codex task

You are rewriting a folder of drafts into one person's writing voice and scanning them for the phrases that mark text as AI-written. Work only from what is given below.

## You are given

- **{VOICE_CARD}** - countable features of the writer's voice: mean sentence length, longest sentence, contractions yes or no, "I" or "we", exact opener and sign-off, questions per 100 words, bullets yes or no, five words they use, five they never use.
- **{DRAFT_FOLDER}** - the folder of drafts to rewrite. Every `.md` and `.txt` file in it, no subfolders.
- **{BANNED_TERMS}** - the tell list: words, phrases and constructions to remove.
- **{SIGNER_NAME}** - the person whose name goes on the finished text.

## Produce

1. `rewritten/<original filename>` for every input file. Same facts, same order of points, voice card applied. Do not add sections, do not add a summary, do not add a closing paragraph.
2. `decoy.txt` in `rewritten/` containing three sentences that each use a different banned term, written by you as a control.
3. `scan.csv` with the columns: `file,term,line,context`. One row per hit, run over `rewritten/` including `decoy.txt`.
4. `stats.csv` with the columns: `file,sentences,mean_sentence_words,longest_sentence_words,contractions,first_person_form`.

## Rules

- Change wording only. Never add, remove or alter a name, number, date, price, URL or quotation. If a draft states a figure, it appears unchanged in the rewrite.
- Never invent a fact, a number or a quote. If a sentence is unclear, leave it as it is and list it in `stats.csv` context.
- Remove a banned term by rewriting the sentence, not by deleting the sentence.
- Every rewritten file ends with a line naming the sender: `Sent by {SIGNER_NAME}`.
- Keep each rewrite within 15 per cent of the source word count.
- Public information only. Do not look anything up and do not add background the draft does not contain.
- The scan must match whole words, case-insensitively, and must include the construction patterns in {BANNED_TERMS}, not only single words.
- Do not touch anything outside {DRAFT_FOLDER} and the `rewritten/` folder you create.

## Return

Print, in this order:

1. The count of input files read and the count of files written.
2. The three `decoy.txt` hits from `scan.csv`, in full. If fewer than three, print `CONTROL FAILED` and stop.
3. The total number of scan hits in the real rewritten files, then every row if there are any.
4. The `stats.csv` table as plain text.
5. Any file you could not rewrite, and the reason in one line.