Your repo needs an employee handbook for AI agents

What it does

repository-harness drops a documentation scaffold into any repo to orient coding agents before they edit code. It creates AGENTS.md and a docs/ hierarchy—product contracts, feature intake rules, architecture boundaries, test matrices, and decision records—so agents know intent, risk, and validation expectations upfront. A Rust CLI and cross-platform install scripts handle the initial setup.

The interesting bit

The project treats the repository itself as the persistent prompt, shifting context engineering out of chat history and into the codebase. Rather than chasing better LLM prompts, it structures the repo as an operating system for agents, complete with risk classification and story-sized work packets. The README explicitly ties the idea to OpenAI’s “harness engineering” concept.

Key highlights

• AGENTS.md serves as a stable agent shim linking local notes and harness docs
• Work flows through feature intake into story packets with explicit behavior-to-proof expectations
• Rust CLI with install scripts that verify checksums for macOS, Linux, and Windows
• Durable decision records in docs/decisions/ meant to survive across agent sessions
• Explicitly at v0: the repo currently contains no product implementation, only the reusable harness framework

Caveats

• Still at v0 with no application implementation or product contract; the repo is purely the scaffolding framework and methodology
• The project is explicitly early-stage and soliciting real-world agent failure cases to refine its patterns

Verdict

Worth a look if you are tired of AI agents confidently refactoring code they do not understand. Skip it if you want a finished tool rather than an early-stage methodology and folder template.