Reviewing Code Core
Use this skill when
- reviewing a PR, diff, branch, or local changes
- auditing a repository for shipped bugs or hidden implementation gaps
- validating that release notes, changelog bullets, docs, and comments match what the code actually ships
- detecting LLM-generated code smells, fake abstractions, placeholders, and unwired functionality
Review stance
Treat code as plausible until verified, not correct until disproven.
Treat docs, release notes, PR text, comments, docstrings, examples, and tests as claims or hints, not proof.
Never emit a defect without exact file and line support.
Never approve changes without stating what was checked.
Workflow
Phase 0 — Reconstruct intent
Extract intended behavior from:
- diff and changed files
- tests
- docs and README edits
- release notes and changelog bullets
- PR description or task brief
- changed exports, routes, commands, config keys, migrations, schemas, and handlers
Convert prose and scattered signals into atomic obligations.
Phase 1 — Summarize actual behavior
Independently summarize what the code actually does.
Do not compare yet.
Phase 2 — Compare obligations vs implementation
For each obligation, classify:
- SUPPORTED
- PARTIALLY_SUPPORTED
- UNSUPPORTED
- CONTRADICTED
- STEALTH_CHANGE
Phase 3 — Run the substance gate
Check for:
- placeholders, TODOs, stubs, no-ops
- unwired functionality
- fake wrappers and dead abstractions
- duplicate generated logic
- tests that assert existence instead of behavior
- comments, release notes, or docstrings that overstate what exists
- polished scaffolding around missing behavior
Phase 4 — Run AI-slop checks
Check for:
- happy-path-only logic
- off-by-one and boundary failures
- stale API usage
- context rot against local file conventions
- unnecessary async or fake abstractions
- generated duplication
- requirement-conflicting plausible implementations
Phase 5 — Emit verdict
Use exact file/line evidence only.
Classify severity first.
Only then propose fixes.
Severity rules
- CRITICAL — security/supply-chain issue, broken shipped feature, data-loss risk, unsupported shipped claim in a production path, or missing auth/validation at a critical boundary
- HIGH — real bug under normal conditions, stealth shipped behavior, materially unsupported docs/release claim, or serious test blind spot on critical behavior
- MEDIUM — bounded correctness or maintainability issue with plausible impact
- LOW — localized quality issue or drift
- INFO — useful note or coverage caveat
AI-specific pattern names
Use these labels when relevant:
- mapping-hallucination
- naming-hallucination
- resource-hallucination
- logic-hallucination
- claim-hallucination
- phantom-dependency
- stale-api
- context-rot
- happy-path-only
- unwired-functionality
Approval rules
Do not approve if any of the following are true:
- a claimed feature lacks structural proof
- a new dependency is unverified
- a critical trust boundary lacks validation
- a route, command, export, schema, or handler is unwired
- the review cannot cite exact file/line support
If approving, explicitly list:
- files reviewed
- obligations checked
- claims validated
- dependencies validated
- trust boundaries inspected
- tests or commands used as evidence
Output format
VERDICT: APPROVED | REJECTED
RISK: CRITICAL | HIGH | MEDIUM | LOW
ISSUES:
- [severity] file:line — issue
FIXES:
- [priority order remediation]
1---2name: reviewing-code-core-23description: Evidence-first code review for correctness bugs, AI-generated code smells, unsupported claims, unwired functionality, dependency hallucinations, and release-to-code mismatches.4---56# Reviewing Code Core78## Use this skill when9- reviewing a PR, diff, branch, or local changes10- auditing a repository for shipped bugs or hidden implementation gaps11- validating that release notes, changelog bullets, docs, and comments match what the code actually ships12- detecting LLM-generated code smells, fake abstractions, placeholders, and unwired functionality1314## Review stance15Treat code as plausible until verified, not correct until disproven.16Treat docs, release notes, PR text, comments, docstrings, examples, and tests as claims or hints, not proof.1718Never emit a defect without exact file and line support.19Never approve changes without stating what was checked.2021## Workflow2223### Phase 0 — Reconstruct intent24Extract intended behavior from:25- diff and changed files26- tests27- docs and README edits28- release notes and changelog bullets29- PR description or task brief30- changed exports, routes, commands, config keys, migrations, schemas, and handlers3132Convert prose and scattered signals into atomic obligations.3334### Phase 1 — Summarize actual behavior35Independently summarize what the code actually does.36Do not compare yet.3738### Phase 2 — Compare obligations vs implementation39For each obligation, classify:40- SUPPORTED41- PARTIALLY_SUPPORTED42- UNSUPPORTED43- CONTRADICTED44- STEALTH_CHANGE4546### Phase 3 — Run the substance gate47Check for:48- placeholders, TODOs, stubs, no-ops49- unwired functionality50- fake wrappers and dead abstractions51- duplicate generated logic52- tests that assert existence instead of behavior53- comments, release notes, or docstrings that overstate what exists54- polished scaffolding around missing behavior5556### Phase 4 — Run AI-slop checks57Check for:58- happy-path-only logic59- off-by-one and boundary failures60- stale API usage61- context rot against local file conventions62- unnecessary async or fake abstractions63- generated duplication64- requirement-conflicting plausible implementations6566### Phase 5 — Emit verdict67Use exact file/line evidence only.68Classify severity first.69Only then propose fixes.7071## Severity rules72- CRITICAL — security/supply-chain issue, broken shipped feature, data-loss risk, unsupported shipped claim in a production path, or missing auth/validation at a critical boundary73- HIGH — real bug under normal conditions, stealth shipped behavior, materially unsupported docs/release claim, or serious test blind spot on critical behavior74- MEDIUM — bounded correctness or maintainability issue with plausible impact75- LOW — localized quality issue or drift76- INFO — useful note or coverage caveat7778## AI-specific pattern names79Use these labels when relevant:80- mapping-hallucination81- naming-hallucination82- resource-hallucination83- logic-hallucination84- claim-hallucination85- phantom-dependency86- stale-api87- context-rot88- happy-path-only89- unwired-functionality9091## Approval rules92Do not approve if any of the following are true:93- a claimed feature lacks structural proof94- a new dependency is unverified95- a critical trust boundary lacks validation96- a route, command, export, schema, or handler is unwired97- the review cannot cite exact file/line support9899If approving, explicitly list:100- files reviewed101- obligations checked102- claims validated103- dependencies validated104- trust boundaries inspected105- tests or commands used as evidence106107## Output format108VERDICT: APPROVED | REJECTED109RISK: CRITICAL | HIGH | MEDIUM | LOW110111ISSUES:112- [severity] file:line — issue113114FIXES:115- [priority order remediation]