A command-line tool that takes a spec file (markdown describing what to build) and produces a structured diff against the current codebase. The idea: write the spec first, let the tool figure out what changed.
The core loop is: parse the spec into structured requirements, embed both spec and relevant source files, call the model, apply the diff. Each step is a pure function — easy to test, easy to swap out.
Current state
Working for single-file changes. Multi-file changes with cross-file dependencies are the next hard problem. The model is good at producing correct diffs but needs help understanding which files are in scope.
Notes
Writing the spec before touching the code makes the unclear parts visible early. That has been the most useful part so far.