Codex skills are not features. They are capabilities that compound.

I keep getting more out of Codex when I treat skills like a product I own. If a skill feels clumsy or fragile, I do not work around it. I tell Codex to refine it, I adjust the scripts, and I keep a note for the next pass.

Skills should evolve the way products do

A lot of great products are the result of long, boring, continuous improvement. Commercial aircraft are safe and reliable because lessons from every incident over the last 75 years get folded into future designs. The same pattern shows up in cars, medicine, and the infrastructure we take for granted.

Skills let us do the same thing with AI agents. Every deficiency you fix becomes part of the next run. Over time, that makes the agent more reliable, faster, and less wasteful with tokens.

The tradeoff is a little upfront work. The upside is a capability that gets easier to use over time instead of harder.

Logs are a gold mine for new skills

Every so often I ask Codex to review logs and suggest new skills. The logs tell you where the friction is. They show repeated tasks, patterns of confusion, and spots where the agent had to improvise.

I want Codex to surface those patterns for me. It is a cheap way to discover what should become a capability.

The tradeoff here is that you will be tempted to turn everything into a skill. I try to stay selective and focus on the tasks that repeat across weeks, not days.

Automations are where this starts to feel powerful

I have started using Codex for more non-coding tasks too. Organizing files. Creating documents. Calling APIs. The more general the tasks, the more valuable it is to have consistent scripts and skills.

What I am most excited about is Automations in the upcoming Codex app for Windows. It is only on Mac right now, but if the Windows version lands in the next few weeks, I want to see how far this can go.

The vision that sticks with me is maintenance. Theo Browne mentioned a use case while talking about the app for T3 Chat: a Codex automation could check for new model releases daily, then update the T3 codebase overnight to support them. That kind of routine upkeep is the perfect job for an agent on a schedule.

The tradeoff is that automation can hide mistakes. You still need guardrails and reviews. But the upside is huge. It turns maintenance into a background process, not a weekly chore.

The point is compounding leverage

The best part is that these ideas stack. Evolving skills reduce friction. Scripts cut cognitive load. Log reviews reveal the next skill. Automations make the whole system run while you sleep.

If you are doing the same thing twice, it is a clue. Turn it into a capability and let it compound.

If you have your own patterns for evolving skills or Automations you are excited about, I would love to hear them.