The Future of AI Agents Just Arrived (/goal for Claude Code & Codex)
Five sections of a great /goal plan, and the prompt to install the skill
What if your AI agent could keep working overnight - and only stop when the goal is actually met?
That's what /goal does. Claude Code and Codex both shipped it recently, and agents can now run unsupervised for hours on end!
The reason it works? You give the agent a verifiable success condition. Then it loops - run, check, run, check - until the condition is met.
And it works best if your plan is shaped right, so give it these elements:
Definition of Done. Write one short sentence that says when the work is finished. No fuzzy words like "clean" or "polished". For example: "the page saves a CSV with the current filters."
Out of Scope. Name at least two things you do NOT want the agent to build. This stops it from wandering off. For example: no mobile version, no login screen.
A Check Each Turn. Pick one way the agent can test if it's on track - a command, a screenshot, or a small test. Something it can see.
A Turn Limit. Use 25 turns for small jobs, 50 for medium, and 100 for big ones. Do not go higher.
I show three levels of /goal in action in this video.
If you want to try the plan-for-goal skill yourself - here's a simple prompt you can send to your agentic platform:
I'm giving you a skill called Plan for Goal. Get the files with:
git clone https://github.com/robonuggets/plan-for-goal
Follow the README to set it up.
Recommend how it would best apply to our setup in plain language,
and ask me a few questions to clarify my intent before you install it.— Jay
P.S: The full Agentic AI Masterclass is inside the RoboNuggets community - takes you from zero to mastery working with AI agents. Come learn with us if that's of interest! → https://www.skool.com/robonuggets