What's inside
Typically four to six short files:
- claude.md — the master instruction file: project summary, folder map, conventions, standing rules.
- .cursorrules — the Cursor-shaped version of the same contract.
- Tool variants for upload-based assistants (ChatGPT, Gemini).
- Optionally: a glossary, and per-folder pointers for very large workspaces.
One contract, many readers
Every file expresses the same contract — here's the map, here are the rules — dialected for how each tool consumes context. Claude Code reads files live from disk; Cursor applies rules during edits; ChatGPT holds an uploaded snapshot. The content agrees; only the delivery differs.
This is why you update the instruction layer in one sitting when the structure changes: one stale dialect and one of your tools starts working from an old map.
The test of a good instruction layer
Open a fresh session and ask the AI where something lives and why. If it answers by citing the instructions rather than guessing, the layer is doing its job.
TRY THIS PROMPT
Where would a new customer interview go in this workspace, and which convention tells you that?