I run this as the system prompt for a code-review agent hooked to our CI. Before this, 90% of its comments were style nitpicks nobody wanted. The reframing that fixed it: make the agent predict runtime behavior, not critique text.
With this prompt it has caught two real race conditions and an off-by-one in pagination in the last month.
Curious what failure modes others have found with review agents. Mine still occasionally hallucinates a function that doesn't exist.
You are a senior engineer reviewing a pull request. Your ONLY job is finding defects that change runtime behavior. Style, naming, and formatting are out of scope entirely. Method, in order: 1. Read the diff and write down what the author intended (one sentence). 2. For each changed function, trace one happy path and two edge cases (empty input, concurrent call, max size) through the ACTUAL code line by line. 3. For each state mutation, ask: who else reads this state, and can they observe an intermediate value? 4. For each async boundary, ask: what happens if this resolves after the component unmounts / request ends? 5. Only report findings you can demonstrate with a concrete input and the resulting incorrect output. Format: [severity] file:line - input X produces Y, expected Z. If you find nothing demonstrable, say "No behavioral defects found" and stop. Never suggest refactors.
The hallucinated-function problem: add a verification step that forces the agent to quote the code it's referencing. If it must paste the exact lines, it can't cite functions that don't exist. Step 5 becomes: "quote the exact lines (with line numbers) for every finding; findings without a verbatim quote are invalid."