theboogieman/ai-work-harness

AI Work Harness is a local setup that adds structure and record keeping around an AI coding assistant, so the assistant behaves more like a careful colleague than a fast but forgetful one. The project was built after its author burned through forty thousand credits in a month of undisciplined AI use, and it exists so that does not happen again. The core idea is simple. All the working rules live in one markdown file that the AI assistant reads. A handful of small, cheap AI agents handle bookkeeping tasks like logging and note taking, so the more expensive, capable AI model can focus on actual thinking instead of chores. A bash script checks that things are in order before a work session starts, and a local git repository quietly records every change so mistakes can be undone. Setting it up has two parts. First, a sixty second demo script proves the underlying scripts work correctly on your computer, without needing an AI assistant connected at all. Second, an installer sets up a dedicated working folder, called an estate, with the folder structure, a local only git repository, and the small AI agents in place. This step needs a Python environment with a specific package installed beforehand, and it takes about ten minutes. Once set up, working on a task with the AI assistant automatically creates records instead of relying on memory. Every task folder keeps its own log, current status, and notes. Every one-off check you run, whether SQL or Python, gets saved to a notebook that acts as an audit trail. Every file change is automatically committed to the local git history. And a validator script refuses to let a new work session start if the previous one left things in a messy, undocumented state. The project is designed for one person working on one task at a time, currently depends on GitHub Copilot's custom agent and hook features, and works best with VS Code, git, and Python 3.12 already installed. It is released under the MIT license.