You define what you want in markdown specs, list tasks as checkboxes, and run a shell script. The AI works through them one by one.
FEatures: context reinitialises every iteration. Instead of accumulating conversation history until the agent gets confused, each loop starts clean. The agent reads the spec, does one task, exits.
Works with Claude Code and OpenCode out of the box. No dependencies beyond the AI CLI tool itself.
Example use case: I needed to process 50 files with the same transformation. Defined the spec once, listed all files as tasks and ran the loop.