coleam00/harness-engineering-demo — reverse-engineered prompt

Build me a small demo repo that shows how to make Claude Code work more like an engineer on a team, not just a chatbot guessing at the code.

Use a simple existing scheduling app as the thing being worked on, with a Python backend, a Next.js frontend, and Postgres. Around it, create a lightweight harness for Claude Code where I can ask it to plan a feature, implement it from the plan, validate the work, and review the diff. The plan and reports should be saved as markdown so each session can hand off to the next one.

Please include guardrails that automatically run linting, tests, and type checks, block finishing when the backend checks fail, and prevent reading real env files or doing dangerous recursive deletes. Add a simple code search helper so the agent can find symbols and references. Also include a small loop script that can rerun Claude sessions until a task is done.