A skill that makes coding agents review their own work

What it does Autoprompt wraps popular coding agents—Claude Code, Codex, OpenCode, and six others—in a structured workflow. You give it a goal, and it spins up subagents to plan, implement, test, review, and sign off before anything lands. The project claims a 45% drop in failures on the Terminal-Bench 2.1 benchmark, though the README admits the extra time and token costs are planning estimates, not measured results.

The interesting bit The architecture deliberately separates coordinators, managers, execution lanes, and independent checks so no single agent plans, approves, and verifies its own output. That is essentially a bureaucratic layer sold as a feature, which feels honest.

Key highlights

• Supports nine coding-agent hosts, from mainstream tools like VS Code Copilot and Codex to newer entrants like Prime Agent and Reasonix.
• Run controls let you cap concurrency (mode=tokensaver limits active subagents to six) or open every lane (mode=wide), with custom model routing available on only four of the nine supported hosts.
• Benchmark evidence is published for Terminal-Bench 2.1, showing 73/89 tasks solved versus 60/89 without the skill.
• Requires Node.js 20+, Python 3.11+, and Bash 4.3+; it is not a pure-JavaScript deployment.

Caveats

• The 3× time and 2× token costs are planning estimates based on user reports, not retained benchmark logs.
• Custom model routing via agents= is unavailable on five of the nine supported providers, which inherit the parent model instead.
• It stops cold for choices that change the result or actions needing your authority, so “autonomous” is a stretch.

Verdict Worth a look if you already live inside Claude Code, Codex, or OpenCode and are tired of watching your agent confidently ship broken code. Skip it if you need fast, cheap iterations or a fully hands-off coding robot.