Skills: teach your coding agent something once

The second time you explain to your coding agent how to publish a post on this site, you have found a skill. A skill is a short written procedure, kept in the project, that the agent reads when a job matches it. Explain once, write it down, never again.

What goes in one

The steps, in order, for one job. “Publishing a post: run the content lint, check the hero image exists, set the review status, open a pull request to this branch.” The commands. The things that went wrong last time. Anything you would say to a new colleague doing it.

It is a recipe, not a rulebook. Rules that apply to everything go in CLAUDE.md, which stays short because it is read every time. A skill is read only when it matches, so it can be as long as the job is.

The fastest way to write one

Do the job with the agent once. Then paste this.

We just did a job we will do again. Write it up as a skill: a short name, one line on when
it applies, then the steps in order with the exact commands, and a section for things that
went wrong this time and how we fixed them. Keep it under a page. Show me before you save
it, and tell me where it will live.

Read it. Cut anything that was specific to today. Save it.

Why this compounds

After a few months you have a folder of procedures shaped by your own corrections, and a familiar task comes out right without you walking it through. That is not the model getting cleverer. It is a filing habit, and it is the same reason Hermes gets more useful the longer you run it.

The first time it goes wrong, add a line to the skill rather than to the chat.