These 3 skills changed my life

In plain English

An AI “skill” is a short written instruction file that an AI agent can pull in when a job calls for it. Theo shows three he uses constantly when working with coding agents. He says he did not invent them in advance.

He noticed himself repeating the same behaviour and wrote it down. One has the main agent ask a second AI model to review its work, then fix what is right and push back on what is not. The other two are about handing a task over completely.

One keeps going until a change is merged. The other stops just before merging and leaves that decision to him.

Three ideas worth keeping

• Write skills from habits you already repeat, and edit them by hand when the agent behaves in a way you dislike.
• A skill’s description should say when to use it, not what it is.
• For tasks he trusts the agent with, he tells it to prove the fix works, check the automated tests, and only contact him if truly blocked or finished.

What it means for your business

The lesson is not the specific tools. Teams that write down how a repeated job should be done, and let AI follow it, spend less time re-explaining and re-checking. Theo reports that in a sample of 32 times he used his hands-off skill, 27 changes merged without him looking again. That is his own usage, not a general benchmark.

What it means for software teams

Theo pairs hands-off AI work with safety nets: automated checks, regular test releases, and the ability to undo mistakes quickly. He says you need to know which kinds of changes the models can verify themselves and which still need a person.

He also covers a practical quirk: in some tools, invoking skills manually can duplicate your prompt in the agent’s context.

What we found, and what we did

We capture repeatable procedures as named skills that our agents load when they need them, instead of retyping clever prompts. Our own platform, Mibble, carries its own skills, for example how to take a change through review and how to add a feature.

These sit alongside a shared playbook we use across all our work. We published the approach in Build Skills, Then Loop Them Into a Super-Agent. Theo’s specific skills are his own and we have not used them.

Worth watching

• 03:46: the review skill and how its description is written.
• 15:12: why he built the hands-off merge skill and what it tells the agent.
• 19:03: his numbers on how often it worked unattended.
• 21:16: the variant that does everything except the final merge.