Let Cursor Cook

A portable AI skill for Codex, Claude Code, and other external coding harnesses that can shell out to Cursor's agent CLI.

Use it when you want the host agent to act like a manager/team lead while Cursor agents handle the technical cooking: codebase research, implementation, command execution, debugging, and code review.

Install

Install with skills.sh:

npx skills add vkongv/let-cursor-cook

Install for specific agents:

npx skills add vkongv/let-cursor-cook -a codex -a cursor -a claude-code

What It Teaches

• Keep the host harness responsible for requirements, planning, user communication, and final verification.
• Keep the host harness in a manager/orchestrator role, not an implementer role.
• Delegate technical work to Cursor CLI agents by default.
• Use explicit Cursor chat IDs for serious orchestration.
• Use isolated Cursor worktrees for risky or parallel implementation.
• Use separate Cursor reviewer agents for non-trivial code changes.
• Send valid review-fix code changes back to Cursor instead of patching them locally.
• Run closed-loop Cursor passes with explicit goals, checks, exit conditions, and iteration caps.
• Keep native subagents available for context hygiene and verification sidecars.
• Escalate to native subagents or local implementation if Cursor fails twice on the same task.

Requirements

• Cursor Agent CLI available as agent or cursor agent
• A host AI harness that can run shell commands

Check Cursor Agent:

agent --help

Recommended implementation command:

agent -p --trust --sandbox enabled --workspace /path/to/repo "Implement the scoped change and run relevant local tests..."

Use -f only when headless Cursor still blocks necessary local commands:

agent -p -f --trust --sandbox enabled --workspace /path/to/repo "Implement the scoped change. Allowed commands: npm test and npm run lint..."

Skill

The skill lives in SKILL.md.