For months I was prompting Cursor/Copilot like: "fix this" or "make this better." Garbage in, garbage out — the model would guess, and I'd spend more time undoing its guesses than if I'd just written the code myself.

What actually changed things was forcing every prompt into 4 parts:

1. Role — tell it what kind of engineer to act as ("senior TypeScript engineer focused on performance")

2. Context — paste the actual code + any constraints, not a summary

3. Task — one specific, scoped ask. Not "improve this," but "remove duplication without changing the public API"

4. Constraints — what NOT to touch, what tests must still pass, style guide if relevant

Example that's saved me real time on debugging:

You are debugging a production issue.
Symptom: [describe bug]
Relevant code: [paste]
What I've already ruled out: [list]
Give me your top 3 hypotheses ranked by likelihood, and the fastest way to test each one.

Asking for ranked hypotheses instead of "fix this" keeps you in the driver's seat instead of the model confidently guessing wrong and you not noticing until prod.

Curious what prompt structures other senior folks here are using for code review specifically — that's the one I still haven't fully cracked.