write-like-me

A Claude Code skill that analyzes how you write and generates a personalized voice profile for writing and revising prose. It uses linguistic and psychological measurement methods rather than subjective descriptions to capture your actual style from a corpus of your own writing.

The same skill supports multiple voices. You can build one from your primary writing register, another from a different register, and another from an author whose style you want to study, and switch between them within the same skill.

HOW IT WORKS

The skill has two modes.

When no voice is built yet, it runs a 7-stage analysis that discovers your corpus, extracts and filters the text, measures approximately 50 stylometric features per register, mines rules from the measurements and any existing style notes, pauses for your review, generates the voice profile, and runs a held-out verification test.

When a voice is already built, it reads your profile and writes or revises prose in that voice. The universal anti-AI baseline (references/00-universal-baseline.md) is always active, on top of whatever your profile specifies.

GETTING STARTED

Step 1. Gather your writing corpus.

Create a folder and put your writing in it. Good sources:

• Export your Claude conversation history: Settings → Data Controls → Export Data, then unzip the file into the folder
• Emails you have written
• Essays, articles, blog posts, papers
• Notes, journals, or any other text in your own voice

Aim for at least 20,000 words. 50,000 or more gives more reliable results.

Step 2. Run the skill.

Open any Claude Code session and run /write-like-me. The skill will ask where your corpus lives, run the analysis, and show you the mined rules for your review before writing anything. Once you confirm, it writes your voice profile and a standalone report to the folder you specify.

Step 3. Write.

From this point on, run /write-like-me whenever you want to write or revise prose. The skill detects that a voice is already built and goes straight to writing mode.

Optional: Build more profiles.

You can have as many profiles as you want, each stored under its own name in the voices/ folder. Some ideas:

• A profile for professional writing
• A profile for email
• A profile for social media
• A profile built from a published author whose style you want to study (the skill treats their patterns as an influence layer, not hard rules)

Name the profile when you invoke the skill and it switches automatically.

WHAT GETS MEASURED

The analysis covers approximately 50 features across four categories.

Lexical:

Type-token ratio (MATTR, window 100), function word frequencies, hapax legomena ratio, distinctive content words against a general-English baseline.

Syntactic:

Sentence length distribution (mean, median, standard deviation, quartiles), comma rate per sentence, em-dash / semicolon / colon / parenthetical rates per 1,000 words, concession rate, sentence-initial word patterns.

Hedging and stance:

Hedging token density (might, perhaps, possibly, roughly, appears to), booster token density (clearly, certainly, definitely, very), first-person singular and plural rates, second-person rate.

Structural:

Paragraph length distribution, heading density, bullet ratio.

Two optional tiers add readability scores (Flesch-Kincaid, Gunning Fog) via textstat, and POS-rhythm, dependency depth, and passive voice rate via spaCy. Academic grounding for all features is in references/03-methodology.md.

WHAT YOU NEED

• Claude Code
• Python 3.8+
• A writing corpus of at least 20,000 words (50,000 or more recommended for reliable distributions)

The draft checker (voice_check.py) needs only Python 3.8+. It has no installs and no dependencies of its own.

The skill accepts writing from any of these sources:

INSTALLATION

Copy the write-like-me/ folder into your Claude Code skills directory:

~/.claude/skills/write-like-me/

Invoke in any Claude Code session:

/write-like-me

EXPORTING YOUR CONVERSATIONS

Claude.ai. Go to Settings → Data Controls → Export Data. You receive a zip file containing your conversation history. The extraction script accepts the exported JSON directly with --format claude.

ChatGPT. Go to Settings → Data Controls → Export Data. Use --format chatgpt.

The script scripts/extract_author_turns.py filters to your turns only using a configurable author marker. The skill shows you 5 random samples and waits for your confirmation before analysis proceeds.

PYTHON DEPENDENCIES

The base analysis runs on Python stdlib only, with no installs required.

Two optional packages add extended features (install via pip install -r scripts/requirements.txt):

OUTPUT

Running the analysis produces seven files per voice profile, plus a standalone report:

voices/<your-name>/
  01-generative.md       # Quantitative targets, structural patterns, exemplars, useful when writing
  02-corrective.md       # Hard bans, soft checks, mechanical scan table, useful after writing
  03-corpus-source.md    # Provenance, what was analyzed, when, with what filters
  claude-ai-skill.md     # Self-contained export for Claude.ai Skills upload
  profile.json            # Rules for the draft checker, in machine-readable form
  goldens.jsonl            # Golden examples for evaluating the checker against this voice
  eval-baseline.json       # Recorded metrics the checker is graded against on future changes

_STYLE-PROFILE-<DATE>.md   # Human-readable stylometric report (place this wherever you like)

The claude-ai-skill.md file is ready to use in Claude.ai immediately after the build completes. To add it:

1. Go to Claude.ai → Settings → Customize → Skills
2. Either upload claude-ai-skill.md directly, or open the file and paste its contents into the instruction field
3. Save. Claude.ai will apply your voice rules in every conversation where the skill is active.

Unlike the Claude Code skill files, claude-ai-skill.md has no external dependencies. All rules, patterns, and exemplars are inlined in a single file.

The standalone report includes:

• Cross-register comparison table (all measured features across all registers)
• Per-register breakdown (sentence length distribution, function word frequencies, punctuation rates, hedging and booster density, pronoun rates)
• Notes on interpretation (which register is the primary target, what zero counts mean, corpus limitations)
• Provenance and re-run instructions

CHECKING A DRAFT

Every voice now ships with a deterministic checker. Run it on any draft:

python scripts/voice_check.py draft.md --profile voices/<your-name>/profile.json

It prints every hit with a line number and a fix, then a verdict. block hits are zero-tolerance rules from your profile and the universal baseline. review hits are patterns with documented exceptions, so you decide. The exit code is 1 on a block, which makes the checker usable as a gate in scripts, hooks, and CI. Without --profile it runs the universal anti-AI baseline only, which works on any text.

EVALUATING THE CHECKER

The checker is itself tested against golden passages, so a rule change cannot silently loosen or over-flag:

python eval/voice_eval.py score --goldens eval/goldens/universal.goldens.jsonl --gate eval/baselines/universal.json

Each voice gets its own goldens.jsonl built from the rule examples you approved during the build. The metrics that matter are block_recall (does it catch what it must) and false_alarm_rate (does it flag what it must not).

UNIVERSAL BASELINE

Every voice inherits a set of zero-tolerance rules for patterns that AI systems produce by default and that human writers do not. These are in references/00-universal-baseline.md and are active regardless of what any individual voice profile specifies.

The baseline removes em dashes as clause separators, stance adverbials (Importantly,, Notably,), filler openers (It is worth noting that), unsupported evaluative adjectives (innovative, robust), and performative verb choices (delve, leverage, foster). These patterns appear in AI-generated text at rates far above any individual human writer's baseline, and removing them is a prerequisite for the voice profile to be meaningful.

FILE STRUCTURE

write-like-me/
  README.md
  SKILL.md                          # Mode detection, routing, 7-stage workflow
  references/
    00-universal-baseline.md        # Anti-AI-tell rules, always active
    01-corpus-discovery.md          # What makes a good corpus; consent guardrail
    02-author-filtering.md          # Filtering exports; preview/confirm step
    03-methodology.md               # Full academic grounding for all features
    04-rule-mining.md               # Three rule sources: files, stats, negative space
    05-exemplar-selection.md        # How to select and annotate exemplars
    06-skill-emission.md            # What goes in each generated file
    07-verification.md              # Held-out generation test protocol
    08-regeneration-and-diff.md     # Re-run protocol when corpus grows
  scripts/
    stylometry.py                   # Feature extraction
    extract_author_turns.py         # Author-turn filter for conversation exports
    generate_report_from_json.py    # Report generator from JSON profiles
    voice_segment.py                # Markdown draft to prose segments, with line numbers
    voice_rules.py                  # Deterministic detectors, parameterized by profile
    voice_profile.py                # Load, validate, and initialize profile.json
    voice_check.py                  # CLI checker, prints hits and a verdict, sets exit code
    requirements.txt                # Optional Python dependencies
  eval/
    voice_eval.py                   # Scores goldens, saves baselines, runs the regression gate
    README.md                       # Golden format and metrics
    goldens/
      universal.goldens.jsonl       # Ships with the repo
    baselines/
      universal.json                # Ships with the repo
  tests/                            # Unit tests, fixtures, and a negative control for the checker
  templates/
    generated-01-generative.md      # Skeleton for write-time guidance
    generated-02-corrective.md      # Skeleton for post-write checklist
    generated-03-corpus-source.md   # Skeleton for provenance file
    generated-claude-ai-skill.md    # Skeleton for the Claude.ai Skills export
    generated-goldens.jsonl         # Skeleton for a voice's goldens file
    profile.schema.json             # Documented shape of profile.json
    style-profile-report.md         # Skeleton for the standalone report
  voices/                           # Your voice profiles live here (not committed to git)
    <your-name>/
      01-generative.md              # Write-time guidance (Claude Code)
      02-corrective.md              # Post-write checklist (Claude Code)
      03-corpus-source.md           # Provenance (Claude Code)
      claude-ai-skill.md            # Self-contained export for Claude.ai Skills
      profile.json                  # Rules for the draft checker
      goldens.jsonl                 # Golden examples for evaluating the checker
      eval-baseline.json            # Recorded metrics for the regression gate

LICENSE

MIT