Skip to main content

review-peer

Review a pull request when requested by a peer. Use when reviewing PRs, providing code review feedback, or analyzing proposed changes. Supports GitHub and GitLab. Pass --triage to summarize a PR and estimate its review effort without reviewing it.

Source facts

Repository
bendrucker/claude
Last source activity
October 20, 2026 at 00:17
Detected SKILL.md language
English
Stars
18
Forks
1

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

File Explorer
15 files

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
review:peer
description
Review a pull request when requested by a peer. Use when reviewing PRs, providing code review feedback, or analyzing proposed changes. Supports GitHub and GitLab. Pass --triage to summarize a PR and estimate its review effort without reviewing it.
argument-hint
<pr-url-or-number> [--triage]
allowed-tools
["Bash(gh:*)","Bash(glab:*)","Bash(git log:*)","Bash(git diff:*)","Bash(git show:*)","Bash(git branch:*)","Bash(git fetch:*)","Bash(git checkout:*)","Bash(git remote:*)","Bash(git rev-parse:*)","Bash(git cat-file:*)","Bash(git status:*)","Bash(jq:*)","mcp__github","WebFetch","Skill(code-review)","Bash(bun ${CLAUDE_SKILL_DIR}/scripts/:*)"]
# Peer Review Assist me in reviewing this PR: $ARGUMENTS ## Context Your own login as the reviewer (see [tone.md](tone.md) for how to address each party). Whichever platform applies resolves. The other reads `unavailable`. - GitHub user: !`gh api graphql -f query='{viewer{login}}' --jq .data.viewer.login 2>/dev/null | grep . || echo "unavailable"` - GitLab user: !`glab api user 2>/dev/null | jq -r .username 2>/dev/null | grep . || echo "unavailable"` ## Arguments - `--triage`: assess the PR for sequencing instead of reviewing it (see [Triage Mode](#triage-mode)). Default: off, which runs the full review workflow below. ## Triage Mode When `--triage` is set, stay read-only and assess the PR for sequencing. Gather just enough to judge scope: the PR body, the diff stat, and the files touched. Then report two things and stop. - What the PR changes, in one line. - The estimated review effort on the same scale step 4 uses for `code-review` (low, medium, high, xhigh), with one-line reasoning. ## Guardrails - **Must** check with me before submitting. Show file comments and review comment. - **Don't** insist on commenting on every PR. Propose approving with no comment if everything looks good. - **Do** match my writing style. You're commenting as me, not a generic AI assistant. - **Do** ask me about ambiguous code rather than guessing. - **Don't** run interpreter one-liners for library introspection. Read the source or fetch the docs. ## Workflow 1. **Research** - Gather context and identify participants (see [research.md](research.md)) 2. **Context** - Determine review context using repository visibility. Private repositories use [corporate](references/corporate.md) defaults. Public repositories use [open-source](references/open-source.md) defaults. Check visibility via the platform API (`gh api repos/OWNER/REPO --jq .visibility` or `glab api projects/ENCODED_PATH | jq .visibility`). If ambiguous, ask me. 3. **Review** - Examine changed files and existing comments 4. **Delegate** - Run `code-review` for code-quality analysis. `code-review` reads the local diff, so run `gh pr checkout` first if not already on the PR branch. Summarize the diff (rough line count, files touched, sensitive areas) and signals from the PR body, propose an effort level with one-line reasoning, and confirm via `AskUserQuestion` before invoking. Skip the call for trivial PRs (docs-only, dep bumps). Effort heuristics: - **low**: docs-only, dep bumps, config tweaks, trivial fixes (<50 lines) - **medium**: typical features or fixes, single module, ~50–500 lines - **high**: large refactors, multi-module, public API or schema changes, ~500–2000 lines - **xhigh**: security-sensitive (auth, payments, data access), breaking changes, migrations, or a change with extreme blast radius 5. **Think** - Evaluate along two axes. Requirement fulfillment: does the change deliver what was asked (see [requirements.md](requirements.md))? Code quality: evaluate against priorities (see [priorities.md](priorities.md)) and smells (see [smells.md](smells.md)), incorporating `code-review` findings. Keep the axes separate so a clean diff does not mask a missed requirement. 6. **Map** - Write the proposed comments to a JSON file (`id`, `path`, `start_line`, `end_line`, `side`, `comment_type`, `content` per comment), then map each to a platform position: ```bash bun ${CLAUDE_SKILL_DIR}/scripts/mapping.ts map --platform <github|gitlab> \ --comments <path> --diff <path> --commit <head-sha> ``` GitLab also needs `--base` and `--start` from the MR `diff_refs`. The CLI runs the in-diff pre-check and returns `{ payloads, dropped }`. An anchor outside a diff hunk lands in `dropped` (GitHub rejects it with `422 "Line could not be resolved"`), so surface those to me and re-anchor rather than losing them. Skip this step when approving with no comments. 7. **Post** - Show me the mapped set, then on my go post as one batch and choose Approve / Comment / Request Changes based on severity. GitHub: a pending review submitted as a batch. GitLab: draft notes published together. See [tone.md](tone.md) for comment style guidelines. ## Service Support This skill assumes GitHub. For GitLab merge requests, load `gitlab:merge-request` for the submission workflow; use `draft-note.ts submit` to publish draft notes with an optional summary and review decision. Comments post through the programmatic path (`mcp__github` / `gh` / `glab`). On follow-up, resolve addressed threads natively on the platform: `review-threads.ts` for GitHub, the resolve flow in `gitlab:merge-request` for GitLab.
View on GitHub