Menu
Launch a multi-lens PR review workflow for blocking issues.
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/nathanvale/claude-code-config --skill pr-review-loopThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Creating or changing tests, fixtures, mocks, snapshots, test helpers, or test harnesses; require a pre-write Test Design Brief.
Browser tasks through Warm Chrome; no Chrome for Testing.
Explicit whole-suite test architecture review, visual candidates, or proof-boundary improvements.
Showing SKILL.md
| name | pr-review-loop |
| description | Launch a multi-lens PR review workflow for blocking issues. |
Generate a workflow script that runs N review lenses concurrently over a package diff, adversarially verifies every finding, and loops until dry — then launches it.
Classify from args and stop at the first matching row. Do NOT launch a workflow before the user confirms.
| Signal | Route |
|---|---|
| No args, empty args, or unrecognized args | STOP — show pre-launch DX menu (see below); do not launch |
| Lens names or lens customisation request (e.g. "new reviewers", "use X instead of Y") | STOP — show pre-launch DX menu with proposed lenses listed; wait for user confirmation before launching |
mode: fast or "fast" in args | Skip menu — launch with defaults immediately |
| Package name or path supplied | Launch with pkg override |
baseRef / branch name supplied | Launch with baseRef override |
resume / run ID supplied | Resume paused run via resumeFromRunId |
Findings at 0 blocking, 0 major | Hand off to /bitbucket |
REQUIRED when no args are supplied. Show this block verbatim — state defaults, then present numbered choices and wait for user input. Do not proceed until the user replies.
Ready to launch PR review.
Defaults:
packages/portal-uiorigin/develop1 to start now.packages/my-pkg).origin/main).Wait for the user to reply before taking any action.
references/workflow-template.jsskills/skill-author/references/skill-io-shape-examples.md#skill-io-example/tmp/pr-review-loop/<runId>/findings.json (written by the generated workflow)All have defaults — pass only what you want to override.
| Parameter | Default | Description |
|---|---|---|
repo | current working dir | Absolute repo root |
pkg | packages/portal-ui | Package path relative to repo root |
baseRef | origin/develop | Git ref to diff against |
lenses | 4 built-in (storybook, correctness, design-system, pr-readiness) | Array of lens configs (see below) |
maxRounds | 3 | Max review rounds before stopping |
maxFindingsPerLens | 3 | Cap per lens per round |
runId | caller must supply | Timestamp string e.g. "2026-06-18T2010" — Date.now() is banned in workflow scripts |
{
key: "correctness", // id prefix e.g. COR
label: "Correctness",
prompt: "You are a Correctness reviewer..." // full reviewer prompt
}
Pass lenses: null to use the 4 built-in lenses from the template unchanged.
references/workflow-template.js — the canonical script source./tmp/pr-review-loop-<runId>.js).Workflow({ script: <generated>, args: { runId } })./workflows).The template assigns models by phase to match agent capability to task difficulty:
| Phase | Model | Why |
|---|---|---|
| Scout (diff, exports, changed-files) | haiku | Shell commands + paste output |
| Review (lens prompts) | sonnet | Structured code review with judgement |
| Verify/refute | inherited (Opus) | Hardest — must read code and reason about false positives |
| Snapshot writers | haiku | mkdir -p && cat > |
| Synthesize | sonnet | Counting + narrative |
| Write-final | haiku | mkdir -p && cat > |
When generating scripts, preserve these tiers from the template. Custom lenses inherit model: "sonnet" from the review loop unless explicitly overridden.
export const meta must be a pure literal — no variables, spreads, or calls.Date.now(), Math.random(), or argless new Date() — caller passes runId.import must not precede export const meta./tmp/pr-review-loop/<runId>/findings.json via a low-effort agent shell call.Workflow({ scriptPath, resumeFromRunId }) — cached agents replay instantly.process global unavailable in workflow scripts — use args instead.import banned at top level — all imports are dynamic inside agent prompts or unavailable; the template uses no ES imports.Workflow({ scriptPath, resumeFromRunId }).TaskStop before resuming or relaunching.runId? Stamp the current time manually (e.g. "2026-06-18T2100") — do not use Date.now().{ key, label, prompt } objects.resumeFromRunId — completed agents return cached results instantly.0 blocking, 0 major? Hand off to /bitbucket to create the PR.