Brand voice guardian

Brand Voice Guardian

Holds every draft to a written voice standard instead of vibes. Reads the brand's voice profile from a local file, scores the draft on a 5-point rubric with quoted evidence, and rewrites only what fails — preserving the author's facts and structure. The voice file is the source of truth; this skill never invents brand rules.

When to Use

• A draft exists (post, thread, newsletter, page copy) and must match an established brand voice.
• The user asks for a tone check, voice score, or "make this sound like us".
• As the final gate after linkedin-content-pack, x-thread-builder, or newsletter-assembler.
• Not for: creating a voice profile from scratch (interview the user and write voice.md first, then run this skill), or fact-checking content (that's deep-research-brief).

Quick Reference

Action	Command / Call
Locate voice profile	Read path from skill config key content.voice_path (default ./voice.md)
Load profile	read_file on the resolved path
Score draft	Apply the 5-point rubric below, 1–5 each, quote evidence
Compute total	python3 -c "print(sum([4,3,5,4,4]))" with your five scores
Rewrite	Fix only dimensions scoring ≤3; keep facts and structure intact

The 5-Point Voice Rubric

Score each 1 (off-brand) to 5 (exemplary), citing a quoted line from the draft as evidence:

1. Vocabulary — uses the brand's preferred words; avoids its banned words/jargon list.
2. Rhythm — sentence length and cadence match the profile (punchy vs. flowing).
3. Stance — attitude and point-of-view strength match (bold, warm, dry, etc.).
4. Formality — contractions, slang, and address ("you"/"we") match the profile.
5. Signature moves — the brand's recurring devices appear (or its anti-patterns don't).

Procedure

1. Precheck — resolve content.voice_path from skill config; fall back to ./voice.md. Confirm the file exists (read_file). If missing or empty, stop: tell the user the expected path and offer to draft a voice.md from 3 samples of their writing. Do not proceed on an imagined voice.
2. Load — read voice.md fully. Extract into working notes: preferred words, banned words, rhythm rules, stance, formality level, signature moves. If the file lacks one of the five dimensions, mark that dimension "not specified — skipped" rather than guessing.
3. Gather — get the draft text and its destination platform (platform norms can legitimately bend formality; note this before scoring).
4. Score — apply the rubric. For each dimension: score 1–5, one quoted line as evidence, one-line reason. Sum with python3 (Quick Reference). ≥20/25 = pass; 15–19 = targeted rewrite; <15 = full rewrite.
5. Rewrite (if needed) — fix only failing dimensions. Never change facts, claims, links, or the argument's order. Keep length within ±10% of the original unless the user asked for cuts.
6. Deliver — output the scorecard, then the rewrite (if any), then a one-line diff summary of what changed and why. Deliver as a draft; never publish or send anything.

Output Template

# Voice Check: [draft title/first line]
**Profile:** [voice_path] | **Platform:** [where this will run]

| Dimension | Score | Evidence | Note |
|---|---|---|---|
| Vocabulary | x/5 | "..." | ... |
| Rhythm | x/5 | "..." | ... |
| Stance | x/5 | "..." | ... |
| Formality | x/5 | "..." | ... |
| Signature moves | x/5 | "..." | ... |
**Total: xx/25 — [PASS / TARGETED REWRITE / FULL REWRITE]**

## Rewrite
[revised draft, or "No rewrite needed"]

## What changed
- [dimension]: [one-line change summary]

Pitfalls

• Voice file missing or stale — scoring against memory of the brand instead of the file. Recovery: step 1 is a hard gate; if the file can't be read this session, stop and say so. If the file contradicts recent published content, flag the conflict to the user instead of silently picking one.
• Rewrite drift — "fixing the voice" quietly changes claims, numbers, or the CTA. Recovery: after rewriting, diff the two versions line-by-line and confirm every fact, link, and number survived; restore anything that didn't.
• Over-sanding — rewriting a 22/25 draft anyway and flattening what made it good. Recovery: respect the thresholds in step 4; a pass ships with the scorecard and zero edits.
• Platform-blind scoring — docking formality points on an X post for contractions the platform expects. Recovery: capture the platform in step 3 and note in the scorecard where platform norms modified the reading of the profile.

Verification

• Voice profile was read from the configured path this session (quote its path in output)
• All five dimensions scored with a direct quote from the draft as evidence
• Total computed via python3, thresholds applied as written
• Rewrite (if any) preserves every fact, number, link, and the CTA
• Output labeled as a draft; nothing was posted or sent