How to use the requesting-code-review agent skill

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

An agent skill is a folder of instructions that an AI agent loads when a task matches the skill's description. This skill is defined in a file called SKILL.md, and it covers one job: asking for a code review at the right moment.

When the skill applies, the agent gathers the git commit range for the work (a base commit and a head commit), writes a short description of what changed, and fills in a review template called code-reviewer.md. It then launches a review sub-agent with that context. The source says the reviewer receives organised context, never the main session's history, so it judges the work itself rather than the reasoning that produced it.

• Reviews are required after each task in subagent-driven development, after completing a significant feature, and before merging to main.
• Reviews are optional but useful when the agent is stuck, before a refactor (to set a baseline), and after fixing a complex bug.
• Feedback is sorted into Critical, Important and Minor issues. Critical issues are fixed immediately, Important issues are fixed before continuing, and Minor issues are logged for later.
• If the reviewer is wrong, the agent is told to push back with technical reasons, show code or tests that prove the point, or ask for clarification.

Install the requesting-code-review skill

Download the ZIP file for requesting-code-review from this page and unzip it. The folder contains SKILL.md and code-reviewer.md. Put that folder in one of these locations, depending on how you want to use it:

• Personal use across projects: place it in ~/.claude/skills/ on your computer.
• One project only: place it in the .claude/skills/ folder inside that project.
• In the Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there.

Check that you have the requirements

The skill depends on git, because the workflow reads commit SHAs such as HEAD and merge-base. Your project therefore needs a git repository with commit history. It also needs a sub-agent or Task tool that can run a general-purpose agent, since the review is launched that way.

The skill runs no scripts and uses no network access itself. Everything it does is carried out through git commands and the sub-agent.

• git installed and a repository with commits
• A sub-agent or Task tool that supports a general-purpose agent type

Trigger the skill with a natural request

You do not need a special command. Ask the agent in plain language, and mention the moment you are at: finishing a task, finishing a feature, or getting ready to merge. The skill's description covers these situations, so a request like these should bring it in.

If you want to be sure, name the skill directly or ask for a review by a separate reviewer rather than a review in the chat.

• 'I finished task 2. Get a code review before we start task 3.'
• 'The feature is done. Have a reviewer check it against the plan before I merge to main.'
• 'Before we merge, send the changes for code review.'
• 'I am stuck on this bug. Get a fresh review of the approach.'

See what the agent does step by step

The agent first collects the commit range. The skill's example uses HEAD~1 as the base and HEAD as the head, or the merge-base with origin/main. It then dispatches the reviewer with four placeholders filled in: what was built, the plan or requirements, the base SHA and the head SHA.

The reviewer is told to work read-only on the checkout. It inspects history with commands such as git diff, git show and git log, and must not change the working tree, index, HEAD or branch state. The template also asks the reviewer to list any behaviour it sets aside as out of scope, so nothing is dropped silently.

• DESCRIPTION: a short summary of what you just finished
• PLAN_OR_REQUIREMENTS: the expected behaviour, for example a task in your plan file
• BASE_SHA: the starting commit
• HEAD_SHA: the ending commit

Worked example: reviewing a finished task

In the source example, the agent has just finished task 2, which added a validation feature. It finds the base commit by searching the log for the task 1 commit, then reads the head commit, and dispatches the reviewer with a description of the two new functions and the plan file that defines the task.

The reviewer returns its strengths, one Important issue (a missing progress indicator) and one Minor issue (a magic number used for the report interval), plus an assessment that work can continue. The agent fixes the Important issue before moving on to task 3.

Safety notes and limits

The skill does not run tests, linters or other checks. It only prompts the agent to launch a review. If you want automated checks, run them separately.

Because the review is read-only by design, the reviewer should not modify your checkout. The template tells it to use a separate temporary worktree if it needs another version of the code. Review the reviewer's output yourself before acting on it, especially when it recommends changes that you think are wrong.

• Runs scripts: no
• Uses network: no
• Not suitable for projects without git history, since the workflow depends on commit SHAs