Learning to Code With AI? Start With AGENTS.md and Skills

Learning to code with AI feels impossible if you didn't start early.

New tool every week. New model every week. If you don't already have the fundamentals, it's hard to even know what to pay attention to, let alone what to learn first. Everyone's shipping demos and dropping jargon, and you're still trying to figure out where to point your attention.

So let me save you some time. If I were starting today, I'd learn two things before anything else: AGENTS.md and skills. They aren't flashy. They're the two that compound.

AGENTS.md: the rules your agent never forgets

AGENTS.md (or CLAUDE.md if you're working in Claude Code) is a plain markdown file that holds the rules your agent has to follow. The agent reads it automatically, every single time you prompt. So instead of re-explaining how you like things done in every message, you write it down once and it sticks.

That's it. It's a place to put the things you'd otherwise repeat forever. "Use tabs, not spaces." "Never touch the migrations folder without asking." "Match the style of the file you're editing." "Always run the tests before you tell me you're done."

Here's the rule of thumb that tells you when something belongs there. If you catch yourself typing "always do this" or "never do that," stop. That's a rule. Put it in AGENTS.md and never type it again. That reflex, hearing yourself repeat an instruction and moving it into the file, is most of the skill.

Skills: stop re-typing the same prompt

The second concept handles the other kind of repetition: not a rule, but a whole task you keep asking for.

If you find yourself typing the same prompt to do the same thing again and again, that's a skill. Build me a React component like this. Write me an article in this exact format. Build me a slide deck with this layout. Same goes for a workflow you run all the time. Mine is a code review: pull the comments from the latest pull request, compare them against the actual code, give me recommendations, and lay them out so I can review what you found quickly. When I'm running that same sequence every few days, it shouldn't live in my head. It should be a skill.

A skill is just a markdown file with three parts: a name, a description of when to invoke it, and a description of how to do the thing. That's the whole format.

And the best part is you don't even have to write it. Tell the agent, "I've been prompting this code review workflow over and over, turn it into a skill." It will write one, and it'll usually describe your workflow pretty closely. Then you read it back. If a piece is off, you just say so in plain English. "That step was a one-off, take it out." And it fixes it. You end up with a tight, reusable version of the thing you were doing by hand.

Why these two first

Put them together and the immediate payoff is that your prompts get shorter and sharper. You stop re-explaining your preferences and re-typing your workflows, and you start pointing at them instead.

But the real win is what it does to how you think. Once these two tools are in your head, you start catching yourself in real time. "I correct this every single time, that goes in AGENTS.md." "I run this same prompt every couple of weeks, let's make it a skill." You're building a mental muscle for noticing your own repetition and handing it to the agent.

That's the loop that actually compounds. Every rule you write and every skill you capture is a piece of your own thinking that you never have to spend energy on again. The people who feel fast with these tools aren't the ones who memorized the newest model's release notes. They're the ones who keep offloading their own judgment into a system that remembers it.

So don't start by chasing every new tool. Start by writing down the things you already keep repeating. Keep offloading your own thinking into the agent, and let it compound from there.