From Vibe Coding to a Method: 5 Steps to Work With AI
Asking the AI and retrying until it seems to work is fine for a prototype, not for real work. Five steps to give it a method, from brief to review.
If you work with AI by asking, trying and asking again until it "seems to work", this pill is for you. Five steps take you from improvising to a method: brief, plan, test, review, diary. You can try the first one in ten minutes, with the tool you already use.
What It Is
Vibe coding is how almost everyone starts: a quick chat, lots of attempts, ship as soon as something works. Working with a method means that before the AI writes a single line, you say what you want, get a plan, decide how you'll check the result, and at the end you verify and take notes.
It applies to a chat like Claude or ChatGPT, and even more to an AI agent, an assistant that reads files, writes code and runs commands on its own. The more the AI does by itself, the more the method matters.
Why It Matters
A quick chat is fine for a prototype or for playing with an idea. It doesn't hold up for real work: a site your customers use, a back-office tool, a script that touches data. There, a mistake costs money, and "do it better" is not a plan.
The craft doesn't disappear, it moves. People who work well with AI define the problem better, give better context and check the output better. That's the difference between teams that use AI and teams that create value with it. AI amplifies design thinking, not improvisation.
An Example, Before and After
Before: you type "make me a nice contact page". The AI invents fields, colors and copy. You fix one piece at a time, it changes things that were fine, and after a while you no longer know what changed or why.
After: you write a brief. Goal: a contact page with name, email and message. For whom: customers arriving from their phones. Don't touch: menu and footer. Done when: the form submits and shows a confirmation. The AI proposes a plan, you cut what isn't needed, it builds, you test on your phone. One round, not ten.
How to Do It
- Write a brief before you ask. Four lines: what you want, for whom, what must not change, how you'll know it's done. It's the most important part, and it's yours.
- Ask for a plan, not code. Have the AI list the steps and the files it will touch. Read, fix, then give the go-ahead. A wrong plan takes a minute to fix; wrong code takes much longer.
- Decide the test first. Write down how you'll check the result: a page to open, a case to try, an automated test if you use them. If you don't know how to check it, you're not ready to have it built.
- Always review. Ask the AI to re-read its own work against the brief and tell you what's missing or what it changed on its own. Then look at the parts that matter yourself.
- Keep a diary. Two lines for every finished task: what changed and what you learned. Next time, give it to the AI as context so it doesn't start from zero.
Ready to copy. Use it as your first message, in Claude, ChatGPT or your coding agent:
Before you write anything, read this brief and reply only with a plan. BRIEF Goal: (what I want, in one sentence) For whom: (who will use it) Don't touch: (what must stay as it is) Done when: (how I check that it works) In the plan I want: 1. the steps in order, one per line; 2. the files or parts you expect to change; 3. how you will check each step; 4. your questions, if anything is unclear. Don't guess. Wait for my OK before starting.
And this one when the work is done:
Re-read what you did against the brief, point by point. Tell me what is done, what is missing and what you changed without being asked. Then write two diary lines: what changed and what is worth remembering next time.
Try It Small
Pick a small change you'd normally make "by feel": changing the text and color of a button on your site, or fixing a formula in a spreadsheet. Write the brief, paste the first prompt and read the plan. In ten minutes you'll see whether the AI understood what you wanted before it writes a line. If the plan makes you say "no, that's not what I meant", the method just saved you a round of fixes.
Watch Out For
- It's not for everything: for a quick test or a throwaway idea, free chat is perfect. The method is for work someone will actually use.
- You can't delegate the brief: if the AI writes it while you don't know what you want, you've only moved the improvising.
- A plan is not a guarantee: still check the result. The AI can follow the plan and get one detail wrong.
- A diary only works if you read it: paste it at the start of the next session, or have your agent read it.
- More steps doesn't mean slower: it feels that way at first. Then you stop redoing the same thing three times.
Next step: if you want your agent to remember things between sessions, read how to give an AI agent memory. If you need to redesign an app that already exists, the method is the same: redesign with AI without breaking the code.
In short: AI amplifies method, not improvisation: a clear brief, a plan and a check are worth more than ten random attempts.
Newsletter
AI that actually works, in your inbox.
New pills, the tools we use and what we learn on real projects. Few emails, only when there is something useful.
