style-fingerprint
GitHub分析用户已发布的3-5篇文档，提取句式、语气、结构等机械特征生成可复用的风格卡片并保存至大脑。解决AI写作不像用户的问题，确保后续技能输出符合用户个人语调和习惯。
触发场景
安装
npx skills add mohitagw15856/pm-claude-skills --skill style-fingerprint -g -y
SKILL.md
Frontmatter
{
"name": "style-fingerprint",
"description": "Study 3-5 documents the user actually shipped and distil a compact style card — so every skill writes in their voice, not the model's. Use when asked to learn my writing style, make outputs sound like me, build a voice profile, or when a user complains AI drafts don't sound like them. Produces a style card (rhythm, register, structure habits, pet phrases, banned moves) saved to the Brain where every other skill reads it."
}
Style Fingerprint Skill
The #1 complaint about AI drafts is "it doesn't sound like me." This skill fixes it at the root: it studies writing the user has actually shipped, extracts the mechanical, imitable features of their voice, and writes a style card into the Brain — after which every brain-aware skill drafts in their register instead of the model's default.
What This Skill Produces
- A style card (~200-300 words, structured) capturing the user's voice as reproducible rules, not adjectives
- Before/after proof: one paragraph rewritten from model-default into the fingerprinted voice, so the user can verify the card works
- The card saved to
brain/knowledge/style.md(with the user's approval), plus a one-line pointer incontext.mdvoice section
Required Inputs
Ask for (if not already provided):
- 3-5 samples the user wrote and shipped — real emails, updates, PRD sections, posts. More samples of the same genre beat variety. Politely reject samples the user merely approved but didn't write — an edited-by-committee doc fingerprints the committee.
- The target register if samples span several (exec formal vs team casual) — or fingerprint each as a named variant
Extraction Method
Analyse mechanics, not impressions. For each dimension, extract a rule an imitator could follow:
| Dimension | What to extract |
|---|---|
| Sentence rhythm | Median sentence length; short-sentence frequency; do they open paragraphs long or punchy? |
| Register & warmth | Contractions? First person singular or plural? Directness of asks ("please could we" vs "let's") |
| Structure habits | Bullets vs prose ratio; headers or none; do they front-load the conclusion (BLUF) or build to it? |
| Signature moves | Recurring phrases, connectors ("net-net", "the short version:"), characteristic openings/closings |
| Emphasis style | Bold? Italics? Caps? Em-dashes vs parentheses? Emoji policy (which ones, where, never)? |
| Numbers | Precision habits ("~40%" vs "42.3%"), units, how they hedge estimates |
| Banned moves | What never appears in their writing (corporate filler, exclamation marks, "I hope this finds you well", passive voice…) — the banned list does more work than the rest combined |
Then verify: rewrite one neutral paragraph in the extracted voice and check it against the samples. If it reads generic, the card is adjectives, not rules — extract harder.
Output Format
Style card: [name / register variant] — fingerprinted [date] from [n] samples
Rhythm: [rules] Register: [rules] Structure: [rules] Signature moves: [phrases/patterns, quoted from samples] Emphasis & numbers: [rules] Never: [the banned list] Calibration line: (one sentence from the samples that is peak them — future skills imitate toward this)
Proof — same paragraph, twice:
[model-default version] [fingerprinted version]
📥 Save to Brain: propose writing this card to brain/knowledge/style.md and adding a voice: see knowledge/style.md pointer in context.md. Show the write, get a yes, then use ../professional-brain/scripts/brain_write.py … --commit. Brain-aware skills read context.md voice on every run — the fingerprint takes effect immediately and everywhere.
Quality Checks
- Every dimension yields a followable rule, not an adjective ("uses 8-14 word sentences, one 3-word sentence per paragraph" — not "punchy")
- Signature moves are quoted verbatim from the samples
- The banned list has at least 4 entries — voice is defined by what's absent
- The proof paragraph is verifiably different from model-default and consistent with the samples
- The card is ≤300 words — a style card that's an essay never gets applied
Anti-Patterns
- Do not fingerprint from fewer than 3 samples — you'd be fingerprinting one mood
- Do not describe voice with adjectives ("professional yet approachable") — extract mechanics
- Do not merge conflicting registers into one mushy card — name variants ("exec", "team") instead
- Do not include the user's confidential content in the card — rules and short quoted phrases only
- Do not overwrite an existing style card silently — diff against it and show what changed
版本历史
- a38bc30 当前 2026-07-05 11:19
