| name | reviewing-prs |
| description | Use when reviewing a pull request or branch holistically across multiple dimensions (correctness, security, codebase alignment, code quality, test quality, performance). Accepts a PR URL or branch name as argument. |
Reviewing PRs
Overview
Announce: "I'm using the reviewing-prs skill to perform a holistic review."
This skill performs 6 parallel specialized reviews of a pull request or branch, then synthesizes findings with confidence scoring and severity tiers. Each review dimension runs as an independent subagent, and results are merged, deduplicated, and presented as a single structured report.
Input Parsing
The user MUST provide either a PR URL or a branch name as argument.
Mode 1: PR URL
Matches pattern: https://github.com/.*/pull/[0-9]+
- Extract
<owner>, <repo>, and <number> from the URL
- Fetch PR metadata and changed files with
mcp__mcp-broker__github_pull_request_read, requesting the diff or file list plus title, body, comments, and reviews when available.
- If review comments are not included in that tool result, note the gap rather than falling back to
gh.
Mode 2: Branch Name
Any input that does not match the PR URL pattern.
- Determine the default branch:
git rev-parse --abbrev-ref origin/HEAD
Fall back to main if the command fails.
- Fetch the diff:
git diff <default-branch>...<branch>
- No PR metadata is available in this mode.
Gather Context
After obtaining the diff:
- Parse the diff to identify changed files (lines starting with
+++ b/ or --- a/)
- Read the full contents of each changed file using the Read tool
- Read the project's CLAUDE.md if it exists
- If in PR URL mode, include the PR title, description, and review comments
- Assemble everything into a single context block for the reviewers
Dispatch Reviewers
Launch 6 Task subagents in parallel — all 6 in a SINGLE message with 6 Task tool calls. Use Task tool (general-purpose) with model: haiku for each.
Read each prompt file from the skill directory at dispatch time. Each agent's prompt is the prompt file content with the full context package appended.
| # | Reviewer | Prompt File |
|---|
| 1 | Bug Hunter | bug-hunter-prompt.md |
| 2 | Security Reviewer | security-reviewer-prompt.md |
| 3 | Codebase Alignment | codebase-alignment-prompt.md |
| 4 | Code Quality | code-quality-prompt.md |
| 5 | Test Quality | test-quality-prompt.md |
| 6 | Performance | performance-reviewer-prompt.md |
Each agent MUST return findings in this format:
FINDINGS:
- <file>:<line> | <severity> | <confidence> | <description>
NO_FINDINGS (if nothing to report)
Where <severity> is one of: blocker, important, suggestion
Where <confidence> is an integer from 0 to 100.
Synthesize
After all 6 agents return:
- Parse each agent's response for
FINDINGS: or NO_FINDINGS
- Filter — drop any finding with confidence below 80
- Deduplicate — if multiple agents flag the same file:line range (within 3 lines), merge them keeping the highest severity and noting all contributing agents
- Group by severity — Blockers > Important > Suggestions
- Determine verdict:
- Ready to Merge — 0 blockers, 0 important
- Needs Attention — 0 blockers, 1+ important
- Needs Work — 1+ blockers
Output Format
Present results using this template. Omit empty severity sections.
## PR Review: <PR title or branch name>
**Verdict: <verdict>** (<N> blockers, <N> important, <N> suggestions)
---
### Blockers
**[<Category>] <Title>** (confidence: <N>)
`<file>:<line>` — <Agent(s)>
<Description>
### Important
**[<Category>] <Title>** (confidence: <N>)
`<file>:<line>` — <Agent(s)>
<Description>
### Suggestions
**[<Category>] <Title>** (confidence: <N>)
`<file>:<line>` — <Agent(s)>
<Description>
---
<N> agents reviewed <N> files. <N> raw findings → <N> surfaced (80+ confidence).
