A good coding-agent instruction file should answer:
How does this repo work?
What should never change?
How do we check a change?
When should the agent stop and ask?
Write it for the next contributor, too.
A coding agent needs more than access to your repo.
Give it:
A clear task
Examples of good code
Commands to verify its work
Explicit limits
The prompt is only part of the setup.
Unpopular opinion: a small, profitable tool that solves one annoying problem is a more interesting business than another “AI platform for everything.”
The interesting question about AI coding isn’t “can it build an app?”
It’s “can someone else maintain that app six months later?”
That’s the demo I want to see.
A company can have a very boring product and genuinely interesting engineering problems.
Worth remembering when every job search starts with the same ten names.