Name the outcome

I start with what should be different when the task is done. For a code change, that includes the affected workflow. A concrete observation is more useful than a general request to improve something.

Supply the relevant context

I name the relevant files and decisions already made. A source is better than a summary copied from memory. Uncertainty needs to remain visible so an assumption does not turn into an apparent fact.

Define boundaries and checks

The task includes the permitted changes and how to check the result. An agent needs a clear boundary for access and publishing. For a change, I ask for the checks actually run and their results.

Turn mistakes into rules

I analyse my own prompts. Recurring problems inform rules and guard hooks. A concrete rule with a checkable trigger is easier to maintain.