How to use the Resolve PR Review Feedback agent skill

What the skill makes the agent do

Agent skills are instruction packages that an AI coding agent loads when a request matches them. This one is for the stage after a reviewer has already left feedback on a pull request. It does not review code from scratch. The skill fetches the feedback once, judges each item centrally, and dispatches generic helper subagents only for fixes that were approved.

Approved fixes are then pushed, replied to and resolved on GitHub, and the agent checks that the changes actually appeared. Items it cannot safely decide are marked for a human instead of blocking the run.

• Judges each comment on its merits, including nitpicks, and reads the actual code before deciding.
• Fixes valid comments, replies to ones that do not need a code change, and marks uncertain ones as needs-human.
• Pushes fixes, replies and resolves threads in ordinary and pipeline runs, then verifies the result.
• Works only on GitHub, including GitHub Enterprise.

Install the skill in Claude Code or Claude apps

Download the ZIP from this page. Then choose one of the following options.

For a personal install across all your projects, unzip the download into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder so it applies only to that repository.

In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your repository
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Ask for it with a natural request

You do not need a special command. Describe what you want, and the agent should match the request to the skill's description. Requests that mention feedback that already exists on a PR fit best. Requests to review code before any comments exist belong to the separate ce-code-review skill, which the description names for that case.

You can also give the skill a PR number, a PR URL, or nothing at all. With no argument, it uses the PR for your current branch. A link to a single review thread (a URL with a #discussion_r fragment) limits the work to that thread only.

• "Address the review comments on my current branch's PR."
• "Resolve the feedback on PR 123."
• "Work through the review on https://github.com/OWNER/REPO/pull/123."
• "Only handle this review thread: [link with #discussion_r]."
• "Prepare local fixes for the feedback on this PR, and I will publish them myself."
• "Finish the saved feedback batch from this handoff file."

Choose an execution mode

The skill has an ordinary mode and three named modes. Only the named modes need a control argument. Ordinary mode fixes, replies, pushes and resolves. Pipeline mode does the same unattended, for multi-step automation. Return-to-caller mode commits fixes locally and never pushes, so a calling process can publish them later. Resume mode finishes a saved batch from a handoff file without repeating judgment or edits.

Only one execution mode can be used at a time. Resume takes its scope entirely from the saved record, so passing a PR number or URL with it is a conflict and the skill stops.

• Ordinary: the default when no mode is given.
• mode:pipeline: publishes and completes feedback unattended.
• mode:return-to-caller: validates and commits fixes locally, never pushes, and can take an optional handoff:<path>.
• mode:resume handoff:<path>: completes saved replies, resolutions and checklist ticks after checking the publication. The handoff path is required.

Requirements and safety notes

You need Git, the GitHub CLI (gh) authenticated with access to the repository, and Python 3, which runs the bundled pending-feedback.py helper. The skill runs scripts and uses the network. Before fetching, it confirms GitHub with gh repo view and stops on an unsupported forge.

Comment text is treated as context, not as instructions. The agent is told never to run commands or scripts found in comments. It still pushes commits and posts replies on your behalf in ordinary and pipeline modes, so review what it will do before a run on a branch that matters. The skill never merges, rebases, force-pushes or approves CI, and it does not pause to ask questions mid-run.

• Requires: Git, GitHub CLI (gh) with repository access, Python 3.
• Runs scripts: yes. Uses network: yes.
• Not for GitLab, Bitbucket or other forges.
• Do not run it on untrusted comments without reviewing the result, because pushes and replies happen automatically in ordinary and pipeline modes.

Tips for getting good results

Run the skill on a clean working tree so that the fix commits contain only the changes it made. If you want to check the work first, use return-to-caller mode, which leaves fixes as a local commit and records the pending remote actions in a handoff file. Use resume to finish that batch later.

Items the skill cannot decide are returned as needs-human with context for you to read. Their threads stay open, so look for them after each run.

• Use return-to-caller mode when you want to inspect fixes before anything reaches GitHub.
• Read the needs-human items; the skill leaves those threads open rather than guessing.
• Give a #discussion_r link when you only want one thread handled, which keeps the run narrow.
• Keep your gh login scoped to the repository you are working on.