# shibbo

> shibbo tracks the habits that mark AI work as machine-made, in four layers: prose, interface, code and agent behavior. It keeps a dated, sourced registry of these tells, maps which models show which ones (prose against human writers on the same topics; pages, code and final reports from the same fixed briefs), and sets up a coding agent to avoid them.

- Every entry cites at least one source: measured studies (Graphite's AI Tells, Kobak et al.), community lists (Wikipedia's Signs of AI writing) and open-source checkers. shibbo gathers them into one registry.
- Tells change with each release. In 2024 "delve" gave away GPT-4; by 2026 most models use it no more than people do, and newer ones have their own habits.
- shibbo is an editor, not a detector: a clean score means a draft avoids known tells and says nothing about who wrote it.
- Install: `npx shibbo check <file>`; `npx shibbo init <harness>` sets up Claude Code, Codex, Cursor, Copilot, Gemini CLI and other harnesses; `npx shibbo mcp` starts an MCP server.

## Docs

- [What shibbo does](docs/): overview and prior art
- [Quick start](docs/guide/quick-start.html)
- [How it fits your harness](docs/guide/how-it-fits.html)
- [Who pays for the runs](docs/guide/support.html): how to help measure a model
- [Data, tests and limits](docs/concepts/data.html): how the checker was evaluated on held-out topics
- [Scoring](docs/concepts/scoring.html)
- [Sources](docs/concepts/sources.html): every source the registry cites
- [CLI reference](docs/reference/cli.html)
- [MCP server](docs/reference/mcp.html)

## Code and data

- [GitHub repository](https://github.com/xvr2e7/shibbo): MIT-licensed code; the registry is YAML in data/tells.yaml and data/tells.d/
- [npm package](https://www.npmjs.com/package/shibbo)