For a security fix, tell the AI exactly which issue to fix, forbid unrelated refactoring, and ask for a test that fails before the change and passes after. A small diff is easier to review, easier to backport, and less likely to hide a changed check in a pile of cleanup.
The over-engineering habit
Anthropic's prompting guidance notes that some models tend to overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that was not requested. Its sample instruction asks the model to keep solutions minimal: no features or refactoring beyond what was asked, no comments or annotations on code it did not change, no error handling for scenarios that cannot happen, and no helper abstractions for one-time operations.
For ordinary features that is a style preference. For security patches it is a review safety property.
Why small diffs matter for security
A reviewer checking a fix needs to answer one question: does this change close the hole without opening another? That is hard when the diff also renames variables and reformats files. Unexpected changes are a place to hide, accidentally or on purpose, as in the AI-written pull request review checklist where scope creep is a finding in itself.
Small fixes are also easier to backport to older release branches.
A prompt for a minimal fix
Fix only this issue: user-supplied sort parameter is concatenated into an SQL query in reports.py.
Scope: change the smallest amount of code needed. Do not refactor, rename, or reformat anything else.
Do not add new dependencies.
Add one test that fails on the current code and passes after the fix.
Then list every file you changed and why.
If the fix needs a larger change, stop and explain instead of making it.
The last line is important. It gives the model a way to report that the real fix is bigger, rather than silently doing the bigger thing.
Check the result
- Compare the file list with what you expected.
- Read the whole diff, including test changes.
- Confirm the test fails on the original code. A test that passes both before and after proves nothing.
- Look for removed validation. A fix that works by deleting a check "that got in the way" is not a fix.
- Search the diff for new packages or changed configuration.
Anthropic also warns against focusing on passing tests and hardcoding, so be wary of a fix that special cases the test input.
When a bigger change is right
Sometimes the minimal patch leaves a pattern that exists in twenty other places. Handle that as a separate, deliberate change: first the narrow fix to close the issue, then a reviewed refactor with its own tests. Mixing them slows the fix.
Quick check
Should the AI also fix similar issues it notices?
Have it list them instead. Fix them in follow up changes so each review stays small.
Sources
Last reviewed: 19 October 2026.
