Staff Level Review

Staff-Level Review

Use this as a read-only engineering review checkpoint. Produce findings and a verdict; do not edit files, push changes, or post PR comments unless the user explicitly switches to a separate review-fix or posting workflow.

If docs/specs/0004-staff-level-review.md exists in the current repo, it is the fuller definition and takes precedence. If it does not exist, this file is self-contained — do not go looking for the spec elsewhere.

Inputs

Review any bounded diff source:

• PR number, URL, branch, or local diff
• provided patch or diff file
• review context bundle
• changed files plus related docs, ADRs, CONTEXT files, and CI/test signals

If the diff source is ambiguous, ask before reviewing.

Preflight

Before reviewing, decide whether to proceed:

proceed: yes | no | needs-info
reason:
required_context:
suggested_next_action:

Stop instead of forcing a low-confidence review when the PR is closed/merged, the diff is too large, requirements are missing, context is insufficient, or extra repositories are needed but not approved.

Review Scope

Review for high-signal issues across:

• correctness and regressions
• architecture and responsibility boundaries
• security and privacy
• performance and unbounded work
• test quality and missing behavior coverage
• maintainability and naming
• API/interface design
• breaking changes and migration risk
• operability, rollout, observability, and rollback
• requirement compliance

Apply a false-positive filter before reporting: skip issues that are pre-existing, covered elsewhere, already verified by tests/CI, intentionally documented, or merely preference without a clear better alternative.

Test Policy

Do not run broad build or test commands by default.

Allowed by default:

• read CI status or provided test output
• read relevant tests
• reason about missing coverage

Run targeted tests, typecheck, or lint only when the command is known, scoped, and allowed by the user or calling workflow.

Severity And Verdict

Use this severity vocabulary:

• BLOCKING: must be fixed before merge
• SUGGESTION: important improvement worth discussing
• QUESTION: intent is unclear and blocks review confidence
• NIT: minor issue that should not block merge

Verdict is driven by findings:

• Request changes: at least one BLOCKING
• Needs clarification: confidence-blocking QUESTION
• Approve with suggestions: no BLOCKING, but meaningful suggestions
• Approve: no blocking or confidence-blocking findings

Output Format

Findings come first. If there are no findings, say so explicitly. Specific, evidence-backed positives may appear only after findings.

# Staff-Level Review

## Findings

[BLOCKING] Short title

- Issue:
- Why it matters:
- Evidence:
- Recommendation:
- Location:

[SUGGESTION] Short title

- Issue:
- Why it matters:
- Evidence:
- Recommendation:
- Location:

[QUESTION] Short title

- Question:
- Why it matters:
- Evidence:
- What would resolve it:
- Location:

[NIT] Short title

- Issue:
- Recommendation:
- Location:

## Verdict

Approve | Approve with suggestions | Request changes | Needs clarification

## Summary

...

## What's Working Well

- ...

## Verification

- Diff source:
- Context inspected:
- Commands run:
- Checks reviewed:
- Not run:

## Review Constraints

- Read-only:
- Extra repos:
- Disallowed paths:
- Sensitivity:

External References

External or legacy staff-review material may be used as reference only after source and license verification. Do not import external material into this repo unless it has been grilled against docs/specs/0004-staff-level-review.md and the required source registry / notices are updated.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: justinmchoi
• Source: justinmchoi/ai-toolkit
• License: MIT

Install and usage instructions live in the source repository linked above.