Skip to main content
cubic skills give your coding agent instructions for reviewing code, handling PR comments, and using your team’s codebase context. Ask for a task in natural language, and your agent can select the relevant skill. You can also name the skill explicitly. Skills run in your coding agent. They use the cubic CLI for local reviews, the GitHub CLI for PR comment workflows, and the MCP server for wikis, codebase scan findings, and review learnings.

​
Before you start

Set up the tools needed for the workflows you want to use:
For local reviews, we recommend connecting ChatGPT (Codex) or Claude Code for access to more capable review models. You can also use cubic’s included model.

​
Install skills

With Node.js and npm installed, run:
The installer detects supported agents from their configuration files in your home directory and installs skills and commands for each one it finds. --skills-only leaves your MCP configuration unchanged; authenticate MCP separately for the workflows that need it. To choose one agent explicitly, add --to. For example, to install for Codex:
Supported targets are claude, codex, cursor, gemini, opencode, droid, and pi. Use --to universal for agents that read ~/.agents/skills; this target is opt-in.
You can also install the cubic plugin from inside Claude Code:
This includes skills, namespaced slash commands, and the MCP connection. Run /mcp, choose cubic, and complete the browser sign-in flow.

​
Check the installation

Check the installer’s summary for your target and any errors. Start a new agent session in your repository, then ask:
With the cubic CLI installed and signed in, the agent should run a local review, group findings by priority, and offer to fix them. If your working tree is clean, the skill reviews your branch against its base.

​
Prompts to start with

​
PR comments

“Check cubic comments” and the comments command both handle review feedback. They wait for the review to finish, inspect unresolved GitHub review threads, apply and verify worthwhile fixes, then commit and push if code changed. They resolve handled threads, including findings confirmed as false positives or already fixed, without posting replies by default. To only read cubic’s open findings, ask your agent to use the MCP tool directly:

​
Review loops

The cubic-loop skill uses PR review when the current branch has an open PR or you request PR mode. Otherwise, it runs local CLI reviews. It commits fixes between iterations and pushes the branch in PR mode. The default limit is five iterations. The loop can stop earlier when no issues remain or when only low-priority P3 findings remain. Check the final summary for remaining findings.

​
Codebase scan findings

You can provide a repository, a scan issue ID, or an exported CSV. Listing findings does not change your code. When you request fixes, the skill verifies each finding against your current checkout before editing. Commits, pushes, PR creation, and scan triage updates require a separate request.

​
Available skills

​
Slash commands

Command names depend on the agent and how you installed cubic: In Codex, use the installed skills by name or describe the task in natural language. Custom prompt slash commands are deprecated. The same naming pattern applies to these commands:

​
Troubleshooting

Start the agent once so it creates its configuration, or specify the target explicitly with --to, for example --to codex.
Check the installer’s summary for the target you use, then restart your agent session. If needed, reinstall for that target with --force:
For local reviews, complete CLI setup. For PR comment workflows, run gh auth status. For wiki, scan, or learning requests, check your agent’s MCP connection and complete OAuth if prompted.
See Local CLI review for copy-paste AGENTS.md and CLAUDE.md workflow instructions.