Skip to content

Guide · Agent skill

How to use the GitHub PR Review agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The github-pr-review skill is a set of instructions and scripts that lets an AI agent review a GitHub pull request and leave inline comments on it. It is built to keep the review short and checkable, and to show you every draft before anything reaches GitHub. This guide explains what the skill does, how to install it, how to trigger it, and what it needs and touches on your machine.

GitHub PR Review (Human-Voiced)google/adk-recipes · Apache-2.0

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip the github-pr-review folder.
  2. 2Place the folder where agents find skills. Put it in ~/.agents/skills/, a path under skills.paths, or ~/.claude/skills/ or a project's .claude/skills/ if you use Claude.
  3. 3Set SKILL_DIR. Point SKILL_DIR at the folder you installed so the script commands resolve.
  4. 4Confirm gh and python3. Make sure gh is authenticated with access to the repository and python3 is installed.
  5. 5Ask for a review. Say something like 'review PR 123 in owner/repo' or paste the pull request link.
  6. 6Read the draft. Check every proposed comment against the anchored line and remove any you would not defend.
  7. 7Approve and let it post. After approval the agent posts the comments individually, paced one at a time.

What the skill makes the agent do

The skill reviews a pull request and posts inline comments written in a plain, human reviewing voice. It is deliberately bounded: it looks for defects a reader can confirm by looking at the anchored line, not for a deep audit. Each comment is meant to be cheap for the PR author to check.

The agent works in stages. It plans the review based on PR size, analyses the changed files (optionally in parallel sub-agents), keeps only findings that can be verified from the anchored line, writes them in a calibrated style, and shows you the full draft. It reads existing comments so a re-review does not repeat what was already said. Comments are posted only after you approve them, one at a time with pacing.

  • Scales the review to 2-20 comments depending on PR size
  • Drops findings that would need the author to derive a fact
  • Reads existing comments on the PR before commenting
  • Posts only after you approve the draft

Install the github-pr-review skill

Download the skill ZIP from this page and unzip it. Put the github-pr-review folder where your agent looks for skills. The skill's own instructions name ~/.agents/skills/, ~/.config/cloudcode/skills/, or any absolute path listed under skills.paths in ~/.config/cloudcode/cloudcode.json. If you use Claude, you can also place it in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder, or upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

Whatever location you choose, set SKILL_DIR to that folder before running the commands in the skill. Every script call is relative to it.

  • Requires the GitHub CLI (gh), authenticated, and python3
  • Optional: pyyaml. If it is missing the skill continues and reports a warning
SKILL_DIR=~/.agents/skills/github-pr-review

Check the requirements before you run it

The skill needs the GitHub CLI (gh) logged in with an account that can access the repository, plus python3. No other dependency is required. The skill runs scripts and uses the network, so it is worth confirming both before your first run.

The skill's Gather step fetches PR metadata with gh and saves existing comments to a temporary file. You can run that same check yourself to confirm access before asking the agent to review.

  • gh installed and authenticated with access to the repository
  • python3 available on your path
  • pyyaml optional; the skill degrades gracefully without it
gh pr view <PR> --repo <owner/name> --json number,title,body,headRefOid,additions,deletions,changedFiles,state

Trigger the skill with natural requests

You do not need a special command. Ask in plain language. The skill is designed to activate on phrases such as 'review this PR', 'review PR 123', pasting a github.com pull link, or asking for comments on a pull request.

Do not use it for local uncommitted changes or a diff against a branch. For those, the skill's instructions point to the separate review skill.

  • 'Review this PR: https://github.com/owner/repo/pull/123'
  • 'Review PR 123 in owner/repo and draft comments for me'
  • 'Give me inline comments on this pull request'

Read and approve the draft before posting

The approval step is the core of the skill. The agent shows you the full set of draft comments before it posts anything. Read each one and remove any you would not defend if challenged. Nothing is posted without your approval, even when the findings look obviously correct.

The skill is built for reviews you have read. Posting a review you have not checked is outside its intended use.

  • Check that each comment points at a fact on the anchored line
  • Remove comments you cannot stand behind
  • Approve the set only when you are willing to sign it

Limits and safety notes

The skill is deliberately shallow. It may read every changed file in full and follow a call or import one hop into another file in the same project. It may not read third-party or dependency source, run the code under review, or construct exploit payloads. Findings that would need that are meant to be dropped, and the skill records what settling them would require.

Because the skill runs scripts and uses the network, review its files before installing. The skill's instructions also say that its voice guidance is designed so comments look hand-typed, and that using it to hide the AI origin of a review from PR authors is not what it is for.

  • Runs scripts and uses the network
  • Only reads changed files and one hop of same-project imports
  • Does not execute the code under review
  • For large PRs it avoids gh pr diff and works from the checked-out tree

Tips for better results

Give the agent the PR number or link and the repository, so it does not have to guess. For a re-review, run the skill again after a bot or colleague has commented, and it will read the existing comments and avoid repeating them.

Expect a short list. A small set of checkable comments is the intended output. If you want a deep audit for yourself, this skill is the wrong tool.

  • Include the repository name with the PR number
  • Re-run on a PR that already has comments to avoid duplicates
  • Edit the drafts before approving rather than rejecting the whole set

What people use it for

  • A maintainer getting a first-pass review with a short list of checkable comments before their own read-through
  • A contributor reviewing a teammate's PR who wants a draft to edit before posting
  • A team re-reviewing a PR after a bot or colleague has already commented, so the new review does not repeat existing points

Frequently asked questions

How do I install the github-pr-review skill?

Download the ZIP from this page and unzip the github-pr-review folder into a skills location your agent reads, such as ~/.agents/skills/ or ~/.claude/skills/. In the Claude apps you can instead upload it under Settings → Capabilities → Skills. Set SKILL_DIR to the folder before running commands.

What does the github-pr-review skill require?

It requires the GitHub CLI (gh), authenticated with access to the repository, and python3. pyyaml is optional; if it is missing, the skill reports a warning and continues.

Is the github-pr-review skill free to use?

The skill is licensed under Apache-2.0. The source material does not state any price, so check the licence and the page you downloaded from for terms.

What should I use instead for local changes or a branch diff?

Use the separate review skill for reviewing local uncommitted changes or a diff against a branch. The github-pr-review skill is for GitHub pull requests only.

Will the skill post comments without asking me?

No. The agent shows you the full draft and posts only after you approve it. Comments are then posted one at a time with pacing.

Why does the skill miss deeper bugs?

It is designed to stop at what a reader could see in the changed files and one hop of same-project imports. It will not read dependency source or run the code, so deeper issues may be missed by design.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you want a short, bounded first-pass review of a GitHub pull request with comments you can check and approve before they post. Skip it if you need a deep audit, want to review local changes or a branch diff, or intend to post reviews you have not read.

Download the skill