Add a structured AGENTS.md and decision log to an existing project so AI coding agents like Cursor or Claude Code start sessions with full context instead of guessing.
Break a large feature request into risk-classified story packets with defined validation expectations before handing work to an AI agent.
Record architectural decisions in a decisions folder so future AI sessions inherit past tradeoffs instead of re-inventing them.
Install the harness scaffold into any repo with a single shell command, safely merging with existing files.
This repository, described in its README as harness-experimental, is a template and toolset for preparing software projects so that AI coding agents can work on them more reliably. The core idea is that most repositories are organized for human developers who already know the codebase, but AI agents typically start each session with only a chat prompt and a snapshot of files. Without additional context, agents tend to edit code before understanding what the product is supposed to do, miss important constraints, and repeat architectural decisions that were already made. The harness addresses this by adding a set of structured documents to a repository. These include an AGENTS.md file that gives an AI agent a stable starting point, a product contract, a feature intake process for classifying how risky a piece of work is, a test matrix that maps expected behaviors to how they should be verified, story packets for breaking large requests into reviewable units, and a decisions folder for recording tradeoffs so future agents inherit them rather than reinventing them. A shell script can install the harness into any existing project directory with a single command. The installer copies the document structure, creates any missing files while leaving existing ones intact in merge mode, and downloads a Rust command-line tool that serves as the main harness interface. The CLI is verified against a checksum during installation. The project is explicitly described as early-stage. There is no application code in the repository itself. The current content is the harness scaffolding only: file templates, the agent operating model, and a small demo that walks through how a product idea moves from a human specification into stories, validation expectations, and implementation work. Contributions the maintainer is looking for include real examples of agent failures caused by missing repository context, harness installs in actual projects, and workflow patterns that work across different AI coding tools.
← magicwand1993 on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.