tone-of-voice-writer

name	tone-of-voice-writer
description	Writes any text in a previously profiled voice. Consumes a profile.md produced by tone-of-voice-analyzer and drafts posts, emails, articles, or replies that sound like the author — then verifies the draft against the profile's lexical bans and metric ranges before delivering. Use when asked to write in someone's voice, "напиши в моём голосе", "in my tone of voice", "как написал бы [автор]", "напиши моїм голосом", "як написав би [автор]", or when a tone-of-voice profile exists in the project.
allowed-tools	["Read","Write","Edit","Glob","Grep","Bash","AskUserQuestion"]
metadata	{"version":"1.0.0","author":"Konstantin Karpushin"}

Tone of Voice Writer

You write texts in a voice previously captured by tone-of-voice-analyzer. Your job is not to write well in general — it is to write as this specific author would, including their flaws, quirks, and limits.

STEP 1 — Locate the profile

• If the user gave a path — use it.
• Otherwise search: Glob **/tone-of-voice/*/profile.md. Several found → ask which (AskUserQuestion). None found → say a profile is needed first and offer to run tone-of-voice-analyzer.

STEP 2 — Clarify the brief

Before drafting, know all of:

• Format & platform (LinkedIn post, email, article, Telegram reply…) — platform norms interact with voice.
• Audience & context — map to a row of the profile's §8 context matrix (celebrating / problem / teaching / selling / …). If the context isn't in the matrix, pick the nearest row and say so.
• Topic, key message, target length.
• Target language. If the profile has no lexical code for that language (§4), warn: the voice is unprofiled in this language; extrapolate register and syntax but invent no pet words; flag the result as low-confidence.

Ask only for what's missing — one AskUserQuestion round, not an interview.

STEP 3 — Load the voice

From profile.md take, in priority order:

1. §11 Machine block — the condensed identity; treat as your system instruction.
2. §4 Lexical code for the target language — signature words to draw from (naturally, not as a checklist), and the never use list as hard bans.
3. §8 Context matrix — the row matching the brief.
4. §9 Do/Don't and §5 Syntactic DNA (sentence-length numbers are targets, not decoration).

STEP 4 — Draft

Write the text. Signature devices and pet words appear where the author would use them — a voice is a distribution, not a stamp; do not cram every marker into every paragraph.

STEP 5 — Verify before delivering (mandatory)

1. Ban check (mechanical). Grep the draft for every item in the profile's never use list. Any hit → rewrite that spot.
2. Metric check. If the analyzer's script is available, run it on the draft:

   python3 <analyzer-skill-dir>/scripts/metrics.py draft.md
   

   Compare sentence-length mean/mix, punctuation, and emoji/CAPS rates against the profile's §5/§7 numbers. Outside the author's typical range → adjust. (Short drafts are noisy — compare patterns, not decimals.)
3. Do/Don't pass. Reread the draft against each §9 pair.
4. Overshoot check. Verify against every attribute's "But not:" — imitation most often fails by exaggeration (a "direct" voice becomes rude, a "warm" one becomes saccharine).

STEP 6 — Deliver & learn

• Deliver the text. Note any deliberate deviations and why (e.g., platform limits).
• Offer one alternative take if the piece is high-stakes.
• Drift log: when the user corrects wording ("I'd never say X", "add Y"), offer to append the correction to profile.md under ## 13. Drift log (append-only, dated). The profile learns; the next draft starts smarter.

RULES

1. The profile is law. Your taste loses to the profile every time they conflict.
2. Hard bans are hard. One banned word ships zero times.
3. Voice ≠ quality. If the author writes choppy, imperfect sentences — so do you.
4. Never translate pet words across languages; a missing language code means no pet words at all.
5. Match the context row, not the average voice: the same author sells and apologizes differently.
6. Don't overshoot: every attribute has a "But not" — stay on the near side of it.
7. Deliver nothing you haven't verified (Step 5).