AI agent skills: turn workflows into systems

You can now build things, make them sound like you, connect your data, and do it safely. This last lesson is about the jump from doing a task once to having a system that runs whenever you want. It also looks at where all of this goes after the course ends.

The whole idea is repeatability. A workflow you run by hand each time is useful. A workflow you save, so it runs on one instruction, is a system. Skills are how you save one.

Skills: save a workflow once

When you catch yourself walking the agent through the same steps for the third or fourth time, that sequence is ready to become a skill. A skill is a small folder holding one instructions file, a SKILL.md, with a short name and a description at the top. Claude Code keeps skills in .claude/skills, Codex in .agents/skills. You describe the workflow once, carefully, and the description says when it should be used, which becomes the trigger. The next time a task matches, the agent reaches for the skill on its own. You can still call it by name when you want to.

Say every week you pull a few numbers, compare them to last week, and write a short summary. The first time, you spell out every step. Once it is a skill, asking for the weekly numbers is enough, and the routine runs the way you defined it. The setup cost is paid once, and the payoff repeats without you having to remember the routine exists. This is where the time actually comes back.

Why a borrowed stack is noise

The common advice points the other way. Someone publishes a hundred marketing skills, you copy the folder, and the agent supposedly does marketing now. It does not hold up, for a plain reason. A hundred skills from someone else is a hundred of someone else’s decisions, made for someone else’s product. You cannot tell which six matter to you, and the ones that do are buried under the ninety-four that do not.

There is a sharper cost underneath. Every line in a skill that repeats what the agent already knows is wasted space and weaker signal. General marketing knowledge is already in the model. What is worth writing down is the part it lacks: your steps, your standards, and the mistakes it keeps making on your own work. A borrowed stack is mostly the first kind, so its output comes back competent and aligned to nothing.

I make the fuller case for one file over a stack of skills in the Build Log.

Skills worth building

The skills worth building are usually the least glamorous ones: the production jobs you already repeat every week. The people who lean on skills hardest say the same, that the scaffolding work, your templates and boilerplate, pays off first. A few of mine, and what each one actually does:

Notice what these have in common. Each is a process you own, not knowledge the agent lacks. The good ones also carry a short note of what it tends to get wrong on the job. That note is the part no downloaded skill can hand you.

One thing to be clear about. These are mine, wired to my sites, my brand, and my books. Copying the exact files hands you my decisions, and they break the moment your setup differs from mine, which is at once. Take the concept instead, and have the agent write your version. Describe the outcome you want, answer the questions it asks, and adjust from there.

When a routine is ready, hand it over like this:

I keep doing this by hand: [describe the routine in your own words].
Turn it into a skill. First interview me: the exact steps, what a good
result looks like, and the mistakes you tend to make on this kind of
task. Then create a new skill folder and save it as, for example,
.claude/skills/weekly-report/SKILL.md for Claude Code or
.agents/skills/weekly-report/SKILL.md for Codex, with a clear
description of when to use it, and flag anything still thin so I
can fill it in.

Answer it the way you would brief a sharp new hire. The interview is where the work is, and the file is where the answers land. What comes out is a skill shaped to your workflow, which is the only shape that earns its place.

Where this goes next

Everything so far has been you and one agent. The ceiling is higher than that. You can point several agents at one big job at the same time, each taking a piece, then pull their work back together. A full competitive teardown or a deep research run goes that way, one agent per competitor or per angle, covering more ground than one session usually can.

That is the course I am building next. Working with several agents at once is where this stops being a quicker way to do small tasks. It becomes a way to do work that would be too slow or fragmented to do by hand. That jump only makes sense once you are comfortable with everything here, which, if you have worked through these lessons, you now are.

For now you have the foundation. An agent in an editor you control, set up with your voice, connected to your data, safe to run, and able to save what works. That is the whole point of calling this Foundations. Go build something with it.