Perfectly Replicate Writing Skills

完美复刻蒸馏写作技能 (Perfect Replicate Distillation Writing Skill)

Clone any author's writing style by deeply analyzing their content across 7+1 dimensions. Output: (A) a portable Writing Style Prompt usable with any LLM, (B) an installable persona sub-skill for ongoing use.

When to Activate

Trigger on requests for:

• "clone this writer's style" / "write in the style of [author]"
• style analysis of provided articles or content
• building a reusable writing persona from samples
• "make my writing sound more like [person]"

Core Principles (品味原则)

1. Features appearing 3+ times in separate articles are real style signals — once-offs are improvisation, not patterns.
2. Preserve contradictions — when the author writes inconsistently across articles, don't smooth it away. Contradictions are valuable.
3. Be specific, not vague — "平均句长12字，30%的句子少于8字" is useful; "喜欢用短句" is not.
4. Negatives define boundaries — what the author never does often defines style better than what they do.
5. No guessing — never invent author background, biography, or intent. Never pass general writing advice as unique stylistic insight.
6. Don't smooth imperfections — awkward phrasings, quirks, and rough edges are identity markers. Preserve them.

Workflow Overview

This skill runs in 4 phases. Execute them sequentially:

• Phase 0: Ingestion — Collect, deduplicate, and validate source material
• Phase 1: 7+1 Dimension Analysis — Analyze each dimension of writing style
• Phase 2: Style Portrait Synthesis — Cross-validate, extract writing genes, build style card
• Phase 3: Dual Outputs — Generate (A) portable style prompt + (B) installable persona sub-skill
• Phase 4: Blind Test — Generate test pieces and score accuracy

---

Phase 0: Ingestion

Accept input via: direct paste, file paths, or conversation context.

Requirements

• Minimum: 3 articles, ≥2,000 total characters (Chinese) or words (English)
• Ideal: ≥10 articles, ≥10,000 total characters/words
• Detect language automatically (Chinese / English / mixed)

Process

1. Deduplicate — remove near-duplicate content (cosine similarity >0.85 on first 200 chars)
2. Clean — strip boilerplate, site headers/footers, publication metadata
3. Count corpus stats — N articles, total chars/words, per-article breakdown, avg length
4. Collect metadata — ask user for: author name, platform, genre (if not provided)

Quality Assessment

• Insufficient (<3 articles / <2,000 chars): warn user, proceed with "incomplete" flag
• Mixed types: separate by genre, ask user which to distill
• Suspected multiple authors: flag inconsistencies, warn about team-written accounts
• Self-distillation: user's own articles — frame identity as "you are yourself," surface unconscious habits

Updating Existing Clone

If a persona sub-skill already exists for this author, read the existing SKILL.md, do a delta-analysis on new articles only, and incrementally update rather than rebuilding from scratch.

---

Phase 1: 7+1 Dimension Analysis

Load and follow the detailed framework in references/7-dimensions-framework.md. The 7 dimensions plus format adaptation:

Group A — Basic Stats (run in parallel)

1. Lexical — Top-30 word frequency, 10-15 "fingerprint words" (unusually frequent vs. general corpus), collocations (2-3 word pairs), neologism/archaism detection
2. Syntactic — Avg sentence length, clause type ratios (simple/compound/complex), punctuation patterns (dash/colon/semicolon frequency), paragraph length distribution
3. Tonal — Colloquialism index, sentence-length volatility (stddev / mean), lexical diversity (TTR or MATTR)

Group B — Structure (run sequentially)

1. Discourse — Opening types (anecdote/data/question/declaration), closing patterns, paragraph length rhythm (short-long-short patterns), logic flow (deductive/inductive/causal)

Group C — Semantic (run in parallel after B)

1. Rhetoric & Expression — Metaphor/simile frequency, parallelism usage, rhetorical questions, concrete vs. abstract ratio
2. Emotion & Attitude — Polarity distribution (positive/negative/neutral), emotional arc across articles, stance markers (hedging/certainty)
3. Content Perspective — Theme clusters, imagery patterns, pronoun density (1st/2nd/3rd person), value-word clusters

Dimension 8 — Format Adaptation (after all 7 above)

• Structural inertia score (how rigidly the author follows genre conventions)
• Capability matching — which formats/genres the author's style naturally fits
• Formula recommendation — 4-point scale per writing formula: "天生适合" / "可以驾驭" / "不太适合" / "冲突"

Output

Structured analysis with quantified observations for each dimension. Reference references/7-dimensions-framework.md for measurement techniques and thresholds.

---

Phase 2: Style Portrait Synthesis

Distillation

Distill each of the 7 dimensions into 3-5 core conclusions each. Prefer quantified statements over vague impressions.

Cross-Validation: Extract "Writing Genes"

Find 3-5 writing genes — features that:

• Are supported by ≥2 independent dimensions
• Are distinctive to this author (not general good writing advice)
• Would make the text feel wrong if removed

Disallowed List

Generate a 5-10 item disallowed list — specific patterns, tics, moves, or word choices the author never uses. These often define style boundaries better than positive features.

Output: Author Writing Style Card

A structured Markdown document containing:

• Author metadata (name, platform, genre, corpus stats)
• Per-dimension findings (3-5 points each)
• Format adaptation recommendations
• Writing genes (3-5 items)
• Disallowed list (5-10 items)

---

Phase 3: Dual Outputs

Output A — Writing Style Prompt

A self-contained prompt usable with any LLM (Claude, ChatGPT, etc.), containing:

1. Identity statement — 50-100 chars, written in the author's voice, describing who they are
2. Writing Iron Laws — 5-8 non-negotiable rules the LLM must follow
3. Style Reference Card — concise per-dimension guidance with specific examples
4. Format Adaptation — which formats/genres this style works for
5. Mode Selection — recommended writing formulas from the 14 available (see references/writing-formulas.md)
6. Disallowed List — what to never do
7. Pre-Writing Mantra — a short prompt-prefix that primes the style

Output B — Persona Sub-Skill

An installable Claude Code skill at ~/.claude/skills/[author-name]-style/SKILL.md:

• Read references/persona-template.md for the skeleton
• Fill all template slots with the style portrait
• Name: [author-name]-style
• Description: "Write in the style of [author name], a [genre] writer on [platform]"
• Include the full style reference card, iron laws, disallowed list, and mode selection
• Support all 14 writing formulas (see references/writing-formulas.md)

After creating, tell the user it's installed and how to invoke it.

---

Phase 4: Blind Test

Generate two 200-300 word test pieces:

1. Default mode — write naturally in the cloned style
2. Formula mode — write using one of the recommended formulas

Compare each against a randomly-selected 200-300 word excerpt from the source corpus (unseen during analysis).

Match Analysis

Score each dimension as a star rating (★☆☆☆☆ to ★★★★★), producing a composite match percentage.

Include specific evidence: "The test piece uses 'X' (present in original: 12 times across 8 articles) but lacks 'Y' (present in original: 7 times) and introduces 'Z' (0 times in original)."

Quality Gate

If composite score <70% for either test piece, flag discrepancies and offer to re-run a refined version.

---

Special Scenarios

Too Little Material

If <3 articles or <2000 chars: warn user, proceed with "incomplete" flag. All outputs flagged as provisional.

Mixed Article Types

If a single author writes across genres (e.g., technical tutorials + personal essays): separate style by genre, ask user which to distill (or all).

Suspected Multiple Authors

If writing style is inconsistent (e.g., some articles use heavy metaphor, others are purely factual): flag for user review, warn about team-written accounts.

Updating Existing Clone

Read existing SKILL.md, delta-analysis on new articles only, incremental update.

Self-Distillation

User's own writing. Frame identity as "you are the author" — help surface unconscious habits the user may not notice.

---

Reference Files

• references/7-dimensions-framework.md — Detailed measurement framework for all 7+1 dimensions
• references/writing-formulas.md — 14 writing formulas with descriptions and when to use each
• references/persona-template.md — Template for generating the persona sub-skill

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zrh091110225
• Source: zrh091110225/perfectly-replicate-writing-skills
• License: MIT

Install and usage instructions live in the source repository linked above.