Here’s an overview of my golden rules for LLMs so far:

Settings

For Claude /output-style is set to concise - make sure this is written to ~/.claude/settings.json.

Rules (AGENTS.md)

This is just a summary

• never install tools on the host - run tooling in docker with the project dir mounted inside
• create a makefile so its easy to rerun commands, keep make help up to date
• TODO.md tracks phases/milestones/tasks, keep it up to date

Skills

Again, just a summary

• a gauntlet skill, that says the main agent is an orchestrator that manages the tasks and dispatches work to a builder and then a critic/fixer loop until no major issues come back
• a comment reviewer to make sure comments don’t lie, invent stuff, track histroy, etc. - they should be short and correct

In Practice

I tend to let the LLM do a decent amount of research and planning before building. When I’m happy with the requirements I have the LLM create an implementation plan and from there I run the “gauntlet” - I should try running the gauntlet in the first phase also.