Calibrate the /review skill by injecting known bugs into throwaway files, running a review, and scoring what gets caught.
Flow
Step 1: Create Test Fixtures
Create a temporary directory tmp/review-benchmark/ with throwaway source files containing deliberate bugs. Each bug has an ID comment (// BUG-01, // BUG-02, etc.) on the line above it for scoring.
Generate files from the bug catalog below. Each file should be realistic enough that the review has context to work with (imports, types, surrounding logic).
Step 2: Stage as a Diff
git checkout -b review-benchmark-tmp
git add tmp/review-benchmark/
git commit -m "tmp: review benchmark fixtures"
Step 3: Run Review
Run /review 3 and capture the output. The review will analyse the diff containing the injected bugs.
Step 4: Score
For each bug in the catalog, check whether the review flagged it:
- Caught — review identified the issue (even if wording differs)
- Missed — review did not mention it
- Partial — review flagged the area but misidentified the problem
Step 5: Report
Format the results:
## Review Benchmark Results — [project name] — [date]
Detection rate: X/Y (Z%)
### By Category
| Category | Caught | Missed | Partial |
|-----------------|--------|--------|---------|
| Logic | ... | ... | ... |
| Security | ... | ... | ... |
| Dead code | ... | ... | ... |
| Complexity | ... | ... | ... |
| Error handling | ... | ... | ... |
| Test quality | ... | ... | ... |
### Missed Bugs
[List each missed bug with its ID, category, and what it was]
### Recommendations
[Which review checks need strengthening based on misses]
Default mode: print the report to the console.
--report mode: create a GitHub issue with the results:
gh issue create \
--title "Review benchmark: X/Y (Z%) — [date]" \
--label "review-benchmark" \
--body "$REPORT"
This mode is designed for cron/scheduled runs. The issue serves as a historical record of detection rates over time and surfaces regressions.
Step 6: Clean Up
git checkout -
git branch -D review-benchmark-tmp
rm -rf tmp/review-benchmark/
Bug Catalog
Logic Bugs
- BUG-01: Off-by-one in array slice (
arr.slice(0, arr.length - 1)drops last element) - BUG-02: Inverted boolean condition (
if (!isValid)whereif (isValid)was intended) - BUG-03: indexOf truthy check (
if (arr.indexOf(x))misses index 0) - BUG-04: Assignment in conditional (
if (x = 5)instead of===) - BUG-05: Wrong comparison operator (
>=instead of>for boundary) - BUG-06: Silent promise (async call without await, result discarded)
Security Bugs
- BUG-07: Hardcoded API key in source
- BUG-08: User input in RegExp without escaping
- BUG-09: String comparison on auth token (timing attack)
- BUG-10: Missing auth check on destructive endpoint
- BUG-11: XSS via unsanitised innerHTML
Dead Code
- BUG-12: Unreachable code after return
- BUG-13: Both if/else branches identical
- BUG-14: Unused function exported but never imported
Complexity
- BUG-15: 5-level nested conditionals
- BUG-16: Triple-nested ternary
- BUG-17: Function with 80+ lines and 12 branches
Error Handling
- BUG-18: Empty catch block swallowing errors
- BUG-19: Generic catch that loses error type
- BUG-20: Missing error handling on fetch call
Test Quality
- BUG-21: Test with no assertions (just renders component)
- BUG-22: Test that only mocks and checks mock was called
- BUG-23: setTimeout in test instead of fake timers
Each benchmark run generates all 23 bugs. The detection rate across categories shows exactly where /review is blind.