Working with an AI agent
The whole skill of building with an AI agent is a small loop you repeat. Learn the loop and one debugging habit, and you can build almost anything — without writing the code yourself.
The mindset
You are not here to memorise commands. You're here to be the director: you know what you want and whether the result is right; the agent does the building. Your value is the idea and the judgement — "yes, that's it" or "no, more like this".
So you need only two things: a clear picture of what "done" looks like, and the small loop below.
The loop
Everything — a website, an automation, a fix — is this same cycle:
- Describe what you want, in plain words. Concrete beats clever.
- Build — the agent does the work and shows you a result (a page, a command, a flow).
- Run / look at the result. Does it do the thing?
- Feed back — paste what you saw (including any error) and say what's wrong or what to change.
- Refine — repeat until it's right. Each pass is cheap.
That's it. Beginners think building is one giant leap; it's actually many tiny, cheap loops. The agent never tires of another pass.
Describing well
The quality of what you get back tracks the quality of what you ask. You don't need jargon — you need to be specific.
Do
- Say what it's for and who uses it.
- Give an example ("like this, but…").
- Describe what "done" looks like.
- Mention limits ("must work on phones", "free only").
- Ask one thing at a time.
Avoid
- "Make it nice" — nice how?
- Ten requests in one paragraph.
- Assuming it can see your screen or files unless you've shared them.
- Guessing technical terms — plain words are fine.
The debugging superpower
Sometimes a command prints a wall of text, or an angry red error. This is not a wall — it's the answer, in disguise. The habit that makes everything click: grab the whole output and paste it back to the agent with "what does this mean?" or "fix this". It will read it far faster than you can.
So you don't have to select long output by hand, learn these:
| Shortcut | What it does |
|---|---|
| Mac Terminal: ⌘ + ⇧ + A, then ⌘ + C | Selects the entire last result and copies it — ready to paste straight back to the AI. This is the one to remember. |
| Ctrl + L | Clears the screen for a clean start before the next step (your history isn't lost). Mac, Linux and Windows. |
| Mac Terminal: ⌘ + K | Clears further, including what you've scrolled past. |
New to the terminal entirely? The Set up your computer page covers opening it and these shortcuts in full.
Good habits
- Read before you confirm anything irreversible. Sending an email, deleting files, paying — pause and check. The agent should ask first; if it didn't, you can.
- Keep secrets secret. Tokens, passwords and keys don't belong in screenshots, public pages or shared chats. Paste them only where they're asked for.
- Save often. If you're working in a project, committing your work (version control) is a perfect undo button — you can always go back to a moment that worked.
- Small steps win. Ship the tiniest working version, then improve it. A live, plain thing beats a perfect, imaginary one.
What's next
Enough theory — go make something real and feel the loop for yourself.