Prompt reviews for senior engineers

Optimize for accepted changes

A terse prompt that makes an agent touch unrelated files, skip tests or break compatibility costs more than it saved. The thing to optimize is a completed change that meets the requirements, not the length of the first message.

Protect the scope boundary

A technical prompt should name the area of the repository, the interfaces that must stay stable, the tests to run, migration assumptions and the report you expect at the end. Keep explicit “do not modify” lines, configuration keys and version numbers even when they look repetitive.

What Typo checks, exactly

• Protected spans (code, paths, identifiers, flags, numbers, quotes, links) are masked and must come back unchanged.
• The counts of negations (not, never, without, n't and more) and strength words (must, only, always, all, every and more) must be equal before and after.
• A rule that would remove more than three quarters of the words is rejected.
• A change is kept only when the estimated token count goes down.
• These are local, rule-based checks. They do not prove that a model will respond the same way.

Level 9 drops articles and linking words. The constraints survive, but the text reads clipped. For specifications, most teams are better served by level 3.

Failure cases worth testing

• A negative instruction is silently dropped.
• An exact endpoint or file name is changed.
• A retry caused by ambiguity costs more than the first prompt saved.
• An error message needed for diagnosis is paraphrased.
• A tool skips a test suite or changes its definition of done.

Compare before you roll out

Run paired tasks on the same repository state and with the same evaluator. Measure acceptance, the number of model calls, provider usage and total cost. Move a team to a higher level only when quality holds. Safe mode keeps originals on each machine for 30 days if you need to look back.