chatlectify

Compile your AI chat history into a writing-style skill.

chatlectify turns your exported AI conversations (Claude, ChatGPT, Gemini) or any corpus of your own writing into a portable SKILL.md file. Drop that file into a Claude skill folder, paste the system_prompt.txt into any LLM, and the model writes like you — not like the factory-default assistant.

It's a small, local CLI. Your chat data never leaves your machine except when it makes the single LLM call that distills the style file.

Why

Every large model speaks the same corporate-cheerful dialect out of the box: "Certainly!", "Great question!", "In conclusion...". If you've spent a year chatting with it, you have a voice and it has no idea. chatlectify mines that voice from your own messages and hands you back a reusable style artifact.

What you get

Running the pipeline on one export produces:

Install

pip install chatlectify

Or from source:

git clone https://github.com/0x1Adi/chatlectify
cd chatlectify
pip install -e ".[dev]"

Requires Python 3.11+.

Auth

The synth step needs one LLM call. chatlectify auto-detects either:

• An API key: ANTHROPIC_API_KEY (default) or OPENAI_API_KEY.
• A local CLI: the claude CLI (default) or codex CLI.

If neither is available, the pipeline still runs and emits a deterministic fallback skill using your extracted features.

Quick start

1. Export your chat history

• Claude — Settings → Privacy → Export data. Unzip; you need conversations.json.
• ChatGPT — Settings → Data Controls → Export. Unzip; you need conversations.json.
• Gemini — Takeout; the .html file works.
• Your own writing — any folder of .txt / .md / .markdown / .rst files.

2. Build the skill

chatlectify all ./conversations.json --out-dir ./skill

That's it. Open ./skill/SKILL.md.

3. Use the skill

With Claude Code / Claude Desktop: Copy the folder into ~/.claude/skills/<skill_name>/.

With any other LLM: Paste ./skill/system_prompt.txt as the system prompt.

CLI

chatlectify ingest <input>    Normalize an export into messages.
chatlectify features <input>  Extract stylometric features only.
chatlectify build <input>     Build SKILL.md (skip benchmark).
chatlectify benchmark ...     Measure fidelity vs. baseline.
chatlectify all <input>       Full pipeline.

Common options:

• --out-dir PATH — where to write artifacts.
• --provider {anthropic,openai} — default anthropic.
• --model MODEL — override the default model.
• --benchmark — opt in to fidelity measurement (makes 2N extra LLM calls).
• --n N — benchmark sample size (default 100).
• --force — bypass quality gates (use at your own risk).

Run chatlectify <cmd> --help for full flags.

Pipeline

  ingest  ->  clean  ->  features  ->  gates  ->  synth  ->  (benchmark)  ->  emit

• ingest — parses export formats (Claude, ChatGPT, Gemini, plaintext) into a normalized Message[].
• clean — strips code fences, URLs, emails; dedupes near-duplicates; flags pastes.
• features — computes 20+ stylometric metrics (lexical, syntactic, structural).
• gates — blocks low-quality runs: <200 messages, <20k characters, >50% pastes.
• synth — asks one LLM call to distill features + exemplars + anti-patterns into a SKILL.md. Falls back to a deterministic template if the call fails or produces invalid output.
• benchmark (optional) — generates N pairs of baseline vs. skill-prompted completions and measures (a) char-n-gram ROC-AUC between your text and each (lower = more indistinguishable from you) and (b) feature-distance reduction.
• emit — writes all artifacts to --out-dir.

Privacy

• Everything runs locally. The only egress is the single synth call (and optional benchmark calls) to your chosen provider.
• Default .gitignore excludes conversations.json, SKILL.md, and all intermediate JSON — your data and voice stay yours.
• Nothing is telemetered; no analytics; no cloud component.

Development

pip install -e ".[dev]"
ruff check src/ tests/
pytest -ra
pytest --cov=chatlectify --cov-report=term-missing

See CONTRIBUTING.md for contribution guidelines and SECURITY.md to report vulnerabilities.

License

MIT — see LICENSE.