Writing code with GPT is really training your prompt skill.
Some requests: if you know the concept and the proper name, one sentence is enough. If you don’t know what it is called, you describe it in circles, then still have to judge whether it understood you.
So I keep wondering: people who don’t program at all — who don’t even know the name of the feature they want, who can’t state the requirement — how do they vibe-code at all?
A post, 2026. This site is written from that question.
Name the thing
Coding with a model does not train your hands. It trains this: fold a vague wish into a proper noun, then hand it over.
Open the glossary →The post already has an answer
Outsiders do not succeed by being allowed to stay vague. When the brief is vague, the model fills in something it has seen — often wrong, often close. People who get somewhere have already found the name by other means: they point at a product, they drop a screenshot, they pick an English word out of a stack trace.
The curiosity is real. Vibe coding with no nouns is rarely smooth. It is usually good enough. Good enough will ship a page that moves. It will not ship permissions, payments, or concurrency.
One sentence, or a loop
| When you know the name | When you can only describe it |
|---|---|
| Debounce search: wait 300ms after the last key. | “Don’t spin while I type. Search when I’ve thought.” |
| Optimistic favorite; roll back on failure. | “It should change now, but don’t break if the net is bad.” |
| A user may only read rows in their workspace. | “Hide other people’s stuff. Changing the URL shouldn’t help.” |
| Create-order is idempotent; double-click is one charge. | “Don’t charge me twice because my hand was fast.” |
What vibe coding is
Vibe coding is a practice: you describe the software in natural language; the model edits the repository, runs commands, and puts the result in front of you. Your job moves from syntax to editing. You edit the diff, the running page, the error — not the mood in your head.
Mood fails where nouns hold. Debounce, skeleton, row-level security, webhook — each word compresses a paragraph of circling. Those words are dense in the training data. Say them, and the model lands on a common implementation. Don’t, and it guesses among neighbours.
The four steps outsiders already take
- 1.0
Point, then speak
“Like a calendar, but only I can see it.” Pointing at a living product borrows its whole noun set. The model has seen that structure.
- 2.0
Throw a picture in
A screenshot is harder than an adjective. Circle the button in red. The model sees a control, not “that little thing.”
- 3.0
Correct the running result
No spec document. Look at this version, name the one wrong place. The loop lives on the artifact, not in imagination.
- 4.0
Keep the counter-question
The model asks, “Do you mean debounce?” — right or wrong, keep the word. That is how an outsider’s glossary actually grows.
With no programming knowledge you can still ship a personal page, a menu, a sign-up sheet. Those nouns already live on every similar site. The model can assemble them with its eyes closed.
The moment you want “only I can edit,” “paid before open,” “do not charge twice,” you are standing at the door of a proper noun. If you can say it, one sentence is enough. If you cannot, it is not that you are bad at talking — the thing already has a name.