Name the outcome
Describe the user-visible behaviour you want and the acceptance criteria. Include one concrete example of success and any important failure condition. Avoid prescribing a large rewrite when a small fix would solve the problem.
Point to the real project
Provide relevant entry points, commands, conventions, and constraints. Let the agent inspect existing code rather than asking it to guess your architecture. Make it distinguish observed facts from assumptions.
Set operational boundaries
Specify which actions are allowed and which need review, especially publication, destructive changes, and external messages. Keep secrets in supported credential storage rather than prompts or repository files.
Review evidence
Ask what changed, why, what was tested, and what remains uncertain. Read the actual diff and test results. A confident summary is not a substitute for a working result.
Find your starting point