The bad-example, good-example pattern
Artifact: one bad/good pair you'll use the next time you prompt for X
1. The pattern
The strongest example pattern is also the simplest:
Here's a bad version. Here's a good version. Here's what changed.
The model gets two concrete data points and the difference between them. That difference is the rule you're trying to teach, made visible without you having to articulate it.
This is the pattern your brain uses naturally to learn taste — seeing the same thing done two ways and recognizing which is better. You can hand the model the same shortcut.
2. Where it works hardest
Bad/good pairs are most useful where quality is a moving target — code reviews, PR descriptions, commit messages, error messages, log lines, test names, documentation. Anywhere "good" is style-shaped, the pair beats either example alone.
It works because the bad example tells the model what not to produce — which is often more constraining than what to produce. Most prompts only state the positive (make it like X). Adding a negative (not like Y, for these reasons) closes off the wrong half of the answer space.
3. PR review comments, two ways
Write helpful PR review comments. Be specific, actionable, and kind.
[diff]The model will produce reasonable comments. They'll be polite and general. They probably won't be specifically helpful in the way your team needs.
Write PR review comments for the diff below. Avoid the unhelpful pattern; match the helpful pattern.
Unhelpful:
> This could be cleaner.
Helpful:
> This function does two things — parsing and validating. If we split parse() and validate(), the test for the validation rule on line 47 becomes a one-liner instead of needing the full payload setup.
Notice: the helpful one names the problem (doing two things), proposes the move (split), and connects it to a test that gets simpler. Match that pattern.
[diff]The pair plus the meta-note (notice...) teaches the shape of helpful — names the problem, proposes the move, connects to consequence. The model can apply that shape across other comments without you having to write a rubric.