Skip to content

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

HumanWriting

A Codex plugin that helps you write in a voice that sounds like you — not a chatbot.

AI-generated text is easy to spot. The vocabulary is too polished, the sentences all land the same way, and there's a certain... scent to it. HumanWriting strips those tells out. It drafts, revises, voice-matches, and coaches writing that reads natural.

Works with whatever LLM you're already using in Codex — no separate backend, API keys, or cloud dependency. There's an optional backend (FastAPI + Ollama/OpenRouter) for running the detection pipeline as a standalone service, but the plugin itself doesn't need it.

Why?

Every AI writing tool I've tried produces the same thing: competent, soulless prose with a predictable rhythm and a fixed set of pet words ("delve," "testament," "underscores"). It's shelf-stable writing — technically correct, instantly recognizable.

HumanWriting is the opposite direction. It:

  • Removes AI tells — em dashes, rule-of-three lists, "Moreover... Furthermore... Additionally..." chains, "it serves as a testament to," and about 30 other patterns
  • Mixes the rhythm up — varies sentence length, breaks up uniform cadences, adds fragments and hedges the way people actually write
  • Matches your voice — give it a sample of your own writing and it adapts the output to your sentence patterns, word choice, and punctuation habits
  • Scores and iterates — paragraphs flagged by the built-in detector get rewritten, re-scored, and re-checked until they pass

Use it for

  • Drafting essays, stories, emails, and applications
  • Revising existing drafts to sound more natural
  • Coaching — get diagnosis, specific edits, and next steps
  • Turning rough notes into something presentable

Installation

Just clone the repo and install it as a local plugin in Codex:

git clone https://github.com/Jayguoo/HumanWriting.git

In Codex, install from the cloned path. The plugin manifest is at .codex-plugin/plugin.json. The HumanWriting skill is at skills/humanwriting/SKILL.md.

That's it. The skill works with whatever model Codex is already using.

Optional: Backend API (detection pipeline)

The repo also includes a standalone FastAPI server that runs the full draft → detect → rewrite → regenerate pipeline. This is not required for the Codex plugin — it's there if you want to use HumanWriting as an API or run the detector locally.

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Requires Ollama with a model (default: qwen2.5:7b-instruct-q5_K_M), or set HW_OR_MODEL for an OpenRouter model.

run.bat

Starts a FastAPI server on localhost:8000 with:

  • POST /generate — draft + detect + rewrite pipeline
  • POST /humanize — rewrites existing text to be less AI-sounding
  • GET /history — past runs stored in SQLite

How it works

Your prompt → Draft → Critique rewrite → Rule-based humanize → LLM humanize → Detection → Regeneration if flagged → Output

The pipeline runs through multiple passes:

  1. Draft — generates text from a prompt
  2. Critique rewrite — a structural pass that looks for common AI patterns and flattens them
  3. Rule-based humanize — swaps AI-favored words, injects contractions, breaks up tricolons, strips em dashes, adds sentence variance, jitters punctuation
  4. LLM humanize — rewrites each paragraph with voice-rotated prompts so the tone shifts naturally
  5. Detection — scores each paragraph with a heuristic detector (or Binoculars if available)
  6. Regeneration — flagged paragraphs get rewritten targeting specific problem sentences
  7. Post-cleanup — final pass for contractions, fillers, and punctuation

The detector scores from 0.0 (very human) to 1.0 (very AI) based on sentence-length variance and AI-tell word density. Anything above 0.6 triggers another pass.

Stack

  • Backend: FastAPI, SQLite, Ollama or OpenRouter
  • Detection: Heuristic rule-based or Binoculars transformer
  • Plugin: Codex skill system

License

MIT

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages