AI Coding Agent Notes.md

gistfile1.txt

## Rules Files

- Always create a rules file that strictly lays out your approach and coding style
- Have a section called something like "Critcal Rules" and a section called "Strong Guidance"

## Testing

- Prove your use cases e2e
- Tell the agent not to DELETE or SKIP, and to never REMOVE assertions
- Unit tests should be minimal and only to isolate logic
- Force the agent to do Test First. I.e. Write failng test that fails because of bug -> confirm test fails for this reason -> fix bug -> confirm test passes

## Skills

- Skills seem like they will become useful and it's worth defining a few
- Skills don't work properly in Claude Code yet, but they will probably get fixed

## Multi-Agent

- Running multiple agents helps in surprising ways when they are able to communicate with each other
- Add rules to explain how to use the communication tool. 
- Tell them to coordinate together, OR make one agent the boss and the others the subordinates
- Install [Too Many Cooks as a VSIX](https://marketplace.visualstudio.com/items?itemName=Nimblesite.too-many-cooks)
- Follow the instructions on that page to install Too Many Cooks as an MCP into your agent (like Claude Code)

## Specs

- Create specs with clear ID headings that can be shared throughout the codebase
- Tie designs, readmes and tests back to the spec IDs
- Avoid numbering in specs, but ensure the IDs are unique
- Keep code, tests and specs in sync
- The best way to is to edit spec -> tests -> code in this order, but sometimes this becomes unrealistic
-