How to use the Copilot PR Autopilot agent skill

What is an agent skill and what does this one do?

An agent skill is a folder of instructions and optional scripts that an AI agent loads when your request matches the skill's purpose. This one, named copilot-pr-autopilot, is built for loop engineering on a pull request.

Once installed, the agent can request a Copilot code review through GraphQL without you mentioning @copilot, triage every open review thread from Copilot, humans and advanced-security, and dispatch parallel fix sub-agents that follow the repo's build, test and lint conventions. It commits per iteration, replies to each thread citing the pushed commit, resolves the ones it handled, and re-triggers review until the latest review is on the current HEAD and no thread is waiting on the agent's reply.

Open threads left for a human are treated as deliberate hand-offs, such as escalated declines or design tradeoffs, not failures. You review and merge the clean PR.

Install the skill from the downloaded ZIP

Download the ZIP for copilot-pr-autopilot from the skill's page on this directory. Unzip it into your personal skills folder at ~/.claude/skills/ to make it available in all projects, or into a project's .claude/skills/ folder to share it with that repository.

You can also upload the skill in the Claude apps. Go to Settings, then Capabilities, then Skills, and upload the ZIP there.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Check the prerequisites before the first run

The skill drives the GitHub CLI and PowerShell, so both must be ready on the machine where the agent runs. The skill's scripts load a shared library that checks the GitHub CLI on start. If gh is missing or not authenticated, the script stops before doing any work and shows an error naming the install command and gh auth login. Surface that message as is and fix the setup.

You need the GitHub CLI installed and authenticated against the target repository, and PowerShell: Windows PowerShell 5.1 or later, or PowerShell 7 or later. Copilot Code Review is the primary use case but is not a hard requirement.

• GitHub CLI (gh) installed and authenticated against the target repo
• PowerShell 5.1+ (powershell.exe) or PowerShell 7+ (pwsh)
• Copilot Code Review enabled on the repo or account (primary use case, not a hard requirement)
• Triage or Write permission on the target repo for the full loop

Understand permissions: full loop or single iteration

The full multi-round loop needs Triage or Write permission on the repo, because the public API for adding Copilot as a reviewer requires it. If you are an external PR author without write access, the trigger step throws a clear error, and the skill switches to single-iteration mode. In that mode the agent addresses all current findings in one pass, then you either click the 🔄 button next to Copilot in the UI or push a substantive commit to re-trigger review.

• Triage or Write: the agent runs the full hands-off loop
• External author: single-iteration mode, then manual re-trigger

Trigger the skill with natural requests

You do not need to remember script names. Ask in plain language and the agent will match your request to the skill. The skill's trigger phrases are the ones listed in its description.

Good phrasings include the examples below. Name the PR number or be on the PR's branch so the agent knows the target.

• "address copilot comments" on PR 123
• "run a copilot review loop" for this PR
• "fix this PR"
• "iterate on copilot feedback"

What happens during a run

Each round runs a fixed sequence. The parent agent requests the review, waits up to about 20 minutes for Copilot to finish, lists open threads, triages them as fix, decline or escalate, fixes in parallel, runs build and tests, commits and pushes, replies and resolves, then checks convergence. If the check says the loop is not finished, it starts another round. When it converges, it runs a one-time cleanup of outdated threads and reports back with the proof, including the head commit and the review it matched.

The skill has a circuit breaker: every tenth round, the agent recaps all prior rounds against the PR's original scope and chooses to continue, revert drifted commits and ship, or hand off to you.

• Request review, then wait for Copilot
• List and triage open threads
• Fix in parallel, then build and test
• Commit, push, reply and resolve
• Check convergence and repeat or finish

Safety notes: scripts and network access

This skill runs scripts and uses the network. It calls GitHub through the gh CLI and GraphQL, and it makes commits and pushes to the PR branch. Because of that, the loop is not read-only. Only use it on branches where automated commits are acceptable, and check the commit history after each run.

The skill is repository-agnostic and runs the repo's own build, test and lint conventions, so those commands run on your machine. Review the scripts in the skill folder before the first run if your policies require it.

• Runs scripts: yes
• Uses network: yes
• Makes commits and pushes to the PR branch
• Do not use it if you cannot accept automated commits on your branch

Tips for better results

Wait until the PR's structure is stable. The skill itself warns that findings churn each round while a PR is still under active design, which wastes effort. Use it when the PR is functionally complete and you want a final correctness pass.

If you want only human reviewer feedback, use a different workflow. The loop is built around Copilot's comments and treats human threads as items to triage alongside them.

• Use it on a functionally complete PR, not one still being designed
• Check the final commit history and the reply on each thread
• Expect open threads for a human to remain; they are intentional hand-offs