Skip to main content

​
Overview

Skills are a standardized format for packaging AI agent capabilities as structured Markdown files (SKILL.md). Each skill describes a capability — its purpose, usage triggers, and the steps the agent should follow to execute it. This gives AI agents a consistent, discoverable, interoperable way to understand and invoke external tools without manual configuration. You can learn more about the open standard at agentskills.io. CodeRabbit Skills are open-source, agent-installable capabilities built on this standard — browse and contribute at github.com/coderabbitai/skills. With these skills, any compatible agent can trigger a full review without explanations on how to run specific console commands.
Skills vs. manual CLI integrationSkills are a standardized, installable capability — a structured way for agents to discover and invoke tools. Manual CLI integration (covered in the agent integration guides) means directly prompting your agent to run coderabbit CLI commands. The two approaches are complementary: skills provide a consistent, zero-configuration trigger mechanism on top of the same underlying CLI.
Skills complement the existing agent integrations documented in the CLI section. You can use skills alongside manual CLI workflows or as a replacement when you want a more natural, language-driven interface.

​
Prerequisites

Before installing skills with coderabbit skills, make sure the CodeRabbit CLI is installed. Authentication is not required to install or update skills; authenticate before using a skill to start a CodeRabbit review. Install the CodeRabbit CLI:
If coderabbit is not available immediately after installation, start a new shell, or run the exact shell-specific reload command printed under Next steps. If the install script offers browser sign-in without first asking for a region, EU users should skip it and authenticate later with an explicit --region eu.
For full CLI setup instructions, see the CLI overview.

​
Supported agents

The CodeRabbit CLI detects and installs user-level skills for Codex, Claude Code, Cursor, Gemini CLI, and GitHub Copilot. The package-runner alternative supports 35+ compatible agents; see the coderabbitai/skills repository for the complete list.
If you prefer to configure agent integrations manually, see the dedicated guides for Claude Code, Cursor, Codex, and Gemini CLI.

​
Install, update, or uninstall skills

Manage CodeRabbit skills for supported agents with the first-party CodeRabbit CLI commands: coderabbit skills without a subcommand runs coderabbit skills install. install and update:
  1. Download the latest published coderabbitai/skills release and verify its release manifest and checksums.
  2. Detect supported coding agents already installed on your machine.
  3. Show every planned install or update, including the exact destination paths and anything they will leave unchanged.
  4. Ask once for confirmation, defaulting to No, before writing any skill files.
The subcommands require CLI 0.9.1 or later. If coderabbit skills --help doesn’t list them, update the CLI with coderabbit update (Homebrew: brew upgrade coderabbit) and check the result with coderabbit --version. CLI 0.7.6 doesn’t apply automatic updates reliably, so update it manually this way. For an organization-managed or self-hosted installation, use your administrator’s update process. Without --agent or --confirm, the commands need an interactive terminal for the confirmation prompt. In a script or agent conversation, use the agent flow.

​
Set up skills from an agent

When the CLI detects that a coding agent is running it, install, update, and uninstall print a preview without writing skill files or ownership state, unless the command includes --confirm <plan-hash> to apply an approved plan. If detection is unavailable or ambiguous, request the preview explicitly with --agent on the subcommand:
The output contains a readable plan followed by an action_required JSON record with status: "awaiting_confirmation" when changes are available. It is not an NDJSON stream. The agent must show the complete plan to the user, including destination paths, replacements or removals of local modifications, and refreshes of externally managed files. Only after the user explicitly approves that plan, the agent runs the exact command from the record’s command field, such as:
Confirmation works without an interactive terminal and applies only to the unchanged plan of the same subcommand: a plan hash from an install preview doesn’t confirm update or uninstall. A changed release, destination, ownership, or file fingerprint requires a new preview and approval. If the user declines, the agent stops without running the confirmation command. Agent detection and a successful preview do not authorize any change. Here, --agent requests an approval preview. The package runner’s skills add --agent <name> option selects an installation target; it is a separate command with separate prompts and ownership.

​
Updates and existing installations

The CLI safely reconciles existing CodeRabbit skills:
  • Missing and outdated CLI-managed skills are offered for installation or update.
  • Locally modified CLI-managed copies can be replaced after the CLI shows the affected paths and you confirm the change.
  • Verified global installs managed by the package runner, and verified Cursor plugin installs, can be refreshed after confirmation. Their original installer retains ownership and may replace them later.
  • Skills from the official CodeRabbit plugins for Codex and Claude Code are never replaced by coderabbit skills; update those through the plugin marketplace.
  • Project-level and generic, unverifiable, ambiguous, duplicate, or cross-owner installations are left unchanged.
coderabbit skills update applies the same rules to skills that are already installed and skips missing ones. If a conflict is left unchanged, resolve it with the owning installer and run coderabbit skills install again.

​
Uninstall skills

coderabbit skills uninstall removes only the skills that the CodeRabbit CLI installed and records as its own. It lists each skill directory it will remove, including locally modified copies, and asks for confirmation, defaulting to No. For copies that were already deleted, it removes only the CLI’s ownership record. Marketplace plugins, skills managed by npx skills or another package runner, project copies, and other externally owned skills are never removed. Remove those with the installer that owns them. CLI-managed copies that can’t be verified safely are left unchanged, and the command exits with code 1.

​
Package-runner alternative

Use a package runner when you need project-scoped installation, agent-specific selection, or an agent outside the five supported directly by the CodeRabbit CLI:

​
Package-runner options

Customize how skills are installed with these flags:
The package runner installs SKILL.md files into ~/.agents/skills or the selected agent’s designated skills directory. Its prompts, flags, and ownership remain separate from coderabbit skills, even when the CLI offers to refresh a verified global package-runner installation.

​
Available skills

​
CodeRabbit review

The CodeRabbit review skill triggers a full CodeRabbit analysis of your current changes directly from an explicit natural-language request.

​
Triggering the skill

Activate the skill using natural language inside your AI agent:
  • “Run a CodeRabbit review”
  • “Review my changes with CodeRabbit”
  • “Use CodeRabbit to check this branch”
  • “Ask CodeRabbit to review my PR”
The agent reads the installed SKILL.md file and automatically invokes the CodeRabbit CLI.

​
How it works

When triggered, the agent will:
  1. Check that the CodeRabbit CLI is installed and authenticated
  2. Run the review on your current changes
  3. Present findings grouped by severity — Critical, Warning, and Info
  4. Optionally fix issues and re-review until the output is clean

​
CLI integration

The skill runs a CodeRabbit CLI review with structured agent output, selects the review scope from your request, and can compare changes against a base branch. The exact CLI flags are an implementation detail and may change independently of the skill’s natural language interface.

​
Autonomous workflow

Skills enable a fully autonomous review-fix loop: Implement → Review → Fix Critical/Warning issues → Re-review until clean The agent continues iterating until no Critical or Warning issues remain, then provides a summary of everything that was fixed.

​
autofix

The autofix skill fetches unresolved CodeRabbit review threads on your current branch’s pull request and helps you apply fixes interactively or in batch.
The autofix skill additionally requires the GitHub CLI (gh) to locate pull requests, query review threads, and post comments. The skill will attempt to install gh automatically if it is not found, but GitHub authentication must be completed manually.

​
Triggering the skill

Activate the skill using natural language inside your AI agent:
  • “Autofix CodeRabbit comments”
  • “Fix CodeRabbit review issues”
  • “Review and fix CodeRabbit feedback”
The agent reads the installed SKILL.md file, finds your open PR, and pulls unresolved CodeRabbit review threads.

​
How it works

When triggered, the agent will:
  1. Check local branch status and warn about uncommitted or unpushed changes
  2. Locate the open pull request for the current branch
  3. Fetch unresolved review threads authored by CodeRabbit
  4. Extract issue metadata and fix prompts from each thread
  5. Let you choose either review-each-fix mode or auto-fix-all mode
  6. Apply selected fixes and create one consolidated commit
  7. Prompt the user if the commit should be pushed upstream

​
CLI commands used by the skill

The skill relies on GitHub CLI plus Git commands:

​
Fix modes

The skill supports two fix paths:
  • Review each issue: See each proposed fix and approve, defer, or modify it.
  • Auto-fix all: Apply all eligible fixes in batch, prioritized from highest severity first.
In both modes, the skill uses CodeRabbit’s “Prompt for AI Agents” (the agent-ready fix instruction) guidance from each unresolved thread as the fix specification.

​
Troubleshooting

Verify that the skill has a SKILL.md file in the directory expected by your agent. Run coderabbit skills again for a supported CLI installation, or repeat the package-runner command that owns the existing installation.
Check two things:
  1. Installation directory: The skill file must be in the path your agent expects (see Supported agents).
  2. Frontmatter validity: The SKILL.md file must have valid frontmatter. Run coderabbit skills to preview whether the CLI can safely update it; unverifiable or ambiguous conflicts are left unchanged.
If you encounter permission errors when writing to the project directory, try installing to your user directory instead:
This error means the CodeRabbit CLI is not installed or not on your PATH. Reinstall the CLI and then restart your shell:
If coderabbit is not available immediately after installation, start a new shell, or run the exact shell-specific reload command printed under Next steps. If the install script offers browser sign-in without first asking for a region, EU users should skip it and authenticate later with an explicit --region eu.
Re-authenticate the CodeRabbit CLI:
For a first login to an EU-hosted CodeRabbit account, use coderabbit auth login --region eu. When no region is specified or saved, the CLI uses the US region.A browser window opens automatically. Sign in to CodeRabbit and the authentication completes in the browser.If you belong to multiple CodeRabbit organizations, the sign-in choice sets your default org only. Review attribution still depends on the current repository. See Organization selection.

​
Verification commands

Use these commands to confirm your CLI setup is working before filing a skills-related issue:
For detailed CLI troubleshooting, see the CLI overview.