Pre-Landing Review

Gate-oriented safety audit for code changes before landing. Uses a structured checklist with two-pass severity triage and blocking/non-blocking classification.

Distinct from pr-review: pr-review is a multi-dimension code quality review. This skill is a gate-oriented safety audit — it uses an external checklist with two-pass severity triage and a blocking/non-blocking classification.

Native alternative: Claude Code's /ultrareview runs a dedicated native review session optimized for bug-finding (Anthropic ships three free per month on Pro/Max plans at Opus 4.7's launch). Use this skill for checklist-driven, gate-oriented blocking classification with a documented triage protocol; use /ultrareview for lightweight bug-hunting on a single change.

Workflow

1. Determine Diff

Identify the changes to review:

• If on a feature branch: diff against the default branch (git symbolic-ref refs/remotes/origin/HEAD)
• If given a PR number: fetch that PR's diff
• If given specific files: review those files

2. Load Checklist

Read references/checklist.md. This is mandatory — if the checklist is unreadable, STOP and report the error.

3. Pass 1 — CRITICAL (blocking)

Review the diff against critical safety categories. These are potential ship-blockers.

SQL & Data Safety

• Raw SQL without parameterization
• Schema changes without migration safety (lock timeout, reversibility)
• Bulk updates/deletes without WHERE clause verification
• Direct column updates bypassing model validations/callbacks

Race Conditions & Concurrency

• Read-then-write without locking
• Unique constraint reliance without database-level enforcement
• Shared mutable state without synchronization
• Queue/background job idempotency

Trust Boundaries

• LLM/AI output used in SQL, shell commands, or rendered HTML without sanitization
• User input reaching privileged operations without validation
• External API responses used without schema validation
• Deserialization of untrusted data

For each CRITICAL finding:

1. Cite exact file and line
2. Explain the specific risk
3. Use AskUserQuestion with exactly three options: Fix now / Acknowledge risk / False positive
4. If "Fix now": make the fix, then re-check
5. If "Acknowledge": record acknowledgment, continue
6. If "False positive": record, continue

4. Pass 2 — INFORMATIONAL (non-blocking)

Review against remaining categories:

Conditional Side Effects — side effects hidden in conditional branches, callbacks triggered by state changes, error handlers silently swallowing failures.

Magic Numbers — unexplained numeric literals, hardcoded thresholds without constants, timeout values without rationale.

Dead Code — unreachable branches, unused imports, commented-out code without explanation.

Test Gaps — new code paths without test coverage, modified behavior without updated tests, missing edge case and error path tests.

Crypto & Entropy — weak random sources for security contexts, hardcoded secrets, missing TLS/encryption for sensitive data in transit.

Time Window Safety — timezone-naive comparisons, daylight saving edge cases, cron expressions not accounting for clock skew.

Type Coercion — implicit type conversions that could lose data, numeric precision loss across boundaries, implicit string encoding at I/O boundaries.

Present all informational findings in a single summary table (file, line, category, description).

5. Gate Classification

• All Pass 1 issues resolved (fixed or acknowledged) → CLEAR TO LAND
• Any unresolved Pass 1 issue → BLOCKED
• Pass 2 issues are advisory — they don't block landing

6. Suppressions

Do NOT flag:

• Test files using test fixtures/factories
• Migration files following framework conventions
• Comments explaining why a pattern is intentional
• Configuration files with documented values
• Type stubs or interface definitions

Output

Gate verdict (CLEAR TO LAND / BLOCKED), critical issues summary with resolution status, informational findings table.

This skill is read-only by default — only modifies code when user explicitly chooses "Fix now" on a critical issue.