Vibe-code your first prototype
You don't need to code. You'll describe a prototype in plain English and let Claude build it on top of a real open-source design system. The skill you're practising is specifying and steering, not programming.
What "vibe coding" means here
Describing what you want, judging what comes back, and steering with feedback — like briefing a developer. Claude writes the HTML; you own the UX decisions.
What you'll produce
One self-contained .html prototype that someone can click through, plus the prompt you used and a short plain-English description of it.
The golden rule
Build only what the main task needs. A prototype is a sketch you can click, not a finished product.
Prompt Studio: design your own prototypes
In Level 1 you followed a recipe. Here you write your own. Start from your own idea, for any kind of app, feature, tool or situation, and build it up with coaching at every step. Then shape the prompt yourself and check its quality before you spend a message on it.
Describe your own scenario on the canvas, with a coaching question, examples and live feedback for each part. Stuck? Let Claude interview you.
Edit the assembled prompt freely, or write one from scratch. Add proven prompt patterns. The quality gate scores it live.
Send it to Claude, judge the result, and steer it with small, precise follow-up prompts.
Same rules as Level 1: one self-contained HTML file and fake data only. What changes is that you decide the scope, and you must be able to explain every decision.
Working on the free Claude plan
The free plan includes Artifacts — the live preview panel where your prototype appears — and you can publish an artifact as a shareable link. The catch is usage: free accounts get a usage budget that refills over a rolling window of a few hours, and big files and long chats use it up fastest. Anthropic doesn't publish a fixed message count, so plan as if every message matters.
✅ Setup checklist
💡 Make your budget last
- Think first, prompt second. Use the prompt builder so your first message is complete.
- One change per message — but batch related tweaks ("fix these three labels").
- Start a fresh chat when a conversation gets long; paste a summary (see the "Handover" prompt).
- Download every working version before asking for changes — you can always go back.
- Hit the limit? Use the wait to click through what you have, or to plan your next prompts on paper.
Privacy: never paste real personal data, names or private information into Claude. Use invented content for the prototype.
Working on Claude Pro
Pro ($20 a month) gives you at least five times the free plan's usage in each rolling 5-hour session, with a separate weekly cap on top, and unlocks unlimited Projects. A Project keeps your rules in one place, so every chat inside it starts already briefed. Anthropic doesn't publish exact message counts, so still plan your prompts.
✅ Setup checklist
💡 What Pro changes
- Projects hold the rules. Format, design system and accessibility rules live in the Project, so each new chat starts already briefed.
- One chat per version. Start a fresh chat in the Project for v2 and v3, and attach the latest HTML. Shorter chats stay faster and more accurate.
- Room to try alternatives. Unsure about a screen? Ask for a second version and compare the two.
- Models: the default model is fine for building. The most capable one uses your allowance faster, so save it for stubborn problems.
- Weekly cap: spread the work across the week rather than one long evening.
Project instructions
Paste this into your Project's instructions. It updates when you choose a design system or fill in the prompt builder.
Privacy: never paste real personal data, names or private information into Claude. Anything in a Project is reused in every chat inside it.
Making open prompting work on the free plan
The free plan includes artifacts and publishing, but usage refills over a rolling few-hour window, and Anthropic doesn't publish exact numbers. Open-ended prompting burns budget fastest, so these habits matter more at Level 2.
Gate before you send
Aim for 8/10 on build prompts. Every missing ingredient tends to cost a follow-up message.
Small builds
Three or four features per build. Prove one flow works, then add the next in a new prompt.
Ask, then build
"Ask me up to 3 questions first" costs one short exchange and often saves a full rebuild.
Fresh chats
Long chats cost more per message. Start a new chat with the latest HTML attached and a short summary.
Download every version
Experiments go wrong. A saved file means "revert" costs nothing.
Out of budget?
Use the wait: click through what you have and draft your next prompts here.
Making open prompting work on Claude Pro
Pro ($20 a month) gives at least five times the free plan's usage in each rolling 5-hour session, with a separate weekly cap, plus unlimited Projects, Research and Claude Code. Anthropic doesn't publish exact numbers. More room doesn't make vague prompts better; it lets good prompts go further.
Projects hold the rules
Put the format, design system and accessibility rules in Project instructions once, instead of repeating them in every prompt.
Keep knowledge current
Re-upload scenario.md when the canvas changes, so Claude builds on your latest thinking.
Spend room on variants
Use the extra usage to build and compare alternatives, not to rebuild a vague first attempt.
Bigger builds, same discipline
Pro handles larger prototypes, but still prove one flow before adding the next. Gate to 8/10 before sending.
Pick the model deliberately
The default model is fine for building. The most capable one uses your allowance faster, so save it for stubborn bugs or tricky logic.
Mind the weekly cap
Spread the work across the week. A marathon evening can leave nothing for later fixes.
Your scenario, from scratch
Any app, feature, tool, service or situation you can imagine. There are no set options here. Work through the canvas: each part has a coaching question, examples if you're stuck, and live feedback. Fill in what you can; every part you complete is one less thing Claude has to guess.
🎤 Prefer to talk? Press the microphone in any field and speak. Say "new line", "comma" or "full stop" for punctuation, and press it again to stop. Your browser may send the audio to its speech service, so don't dictate personal details.
🗣 Stuck before you start? Let Claude interview you
Start a new chat inside your Project and paste this in with your rough idea, even a vague one. Claude asks one question at a time, then writes a summary you can paste back here. Pro has room for a proper interview, so let it push you.
Copy this into a new chat with your rough idea, even a vague one. Claude asks one question at a time, then writes a summary you can paste back here. It uses several short messages, so keep it in its own chat.
📋 Paste Claude's summary to fill the canvas
Paste the interview summary (or any notes with headings like Users: and Main task:). Matching parts fill in automatically; you can edit everything afterwards.
Choose an open-source design system
A design system gives your prototype real, tested components — buttons, forms, alerts — so users react to your design decisions, not to wonky AI styling. Select one and it's written into your prompt automatically.
Prompt builder
Fill this in and you'll get a complete first prompt that asks Claude for a single HTML file with inline CSS and JavaScript, using your design system. The better your answers, the fewer messages you'll spend fixing things.
🎤 Prefer to talk? Press the microphone in any field and speak. Say "new line", "comma" or "full stop" for punctuation, and press it again to stop. Your browser may send the audio to its speech service, so don't dictate personal details.
Prompt library
Follow-up prompts for after your first build. Replace the [square brackets]. Your selected design system is filled in automatically.
Build settings
Choose the device and fidelity, then assemble your prompt from the canvas.
Design system: none yet (chosen in the section above)
Set up your Level 2 Project
On Pro, keep each prototype in its own Project. The instructions below hold the studio's rules, and your canvas becomes a knowledge file. Every chat in the Project, whether interview, build, variants or fixes, then starts already briefed, so your prompts stay short.
1 · Project instructions
Paste into your Project's instructions. It updates as you change the canvas and design system.
2 · Knowledge files
Upload these to the Project's knowledge, and re-upload whenever they change.
3 · One chat per job
- Interview: sharpen the scenario, then update the canvas.
- Build v1: paste your assembled prompt.
- Variants: use the "Three variants" pattern on your riskiest screen.
- Fixes: attach the latest HTML and describe one change at a time.
Shorter, focused chats are faster and more accurate than one long one, and they use less of your allowance per message.
Workspace and quality gate
This is your prompt; edit anything. Assemble it from your canvas, or write it entirely yourself. The gate checks for the ingredients that save you messages later. It never blocks you: a 6/10 prompt for a quick experiment is fine if you know what you left out.
Quality gate
Prompt patterns
Reusable shapes for prompts, not fixed text. Insert one into your workspace and replace the [square brackets]. The illustrated guide shows each one as a diagram.
Fake it honestly
A single HTML file can't do everything a real app does. It doesn't need to; it only needs to feel real enough for the main task. Here's how to simulate the hard parts, and what to tell Claude.
| Real thing | How to fake it | Say to Claude |
|---|
Ethics line: a prototype must never collect real passwords, card numbers or personal data, even by accident. Label fake payment and login screens as "Prototype — don't enter real details".
Troubleshooting
The preview looks plain or unstyled
Claude's preview can block some external files. Download the HTML and open it in Chrome or Edge — if it looks right there, the prototype is fine. Or ask Claude: "Load the design system from cdnjs.cloudflare.com if it's available there."
Claude rebuilt everything when I asked for one change
Say what must not change: "Change only the button text on the Confirm screen. Keep everything else exactly as it is." If it still drifts, go back to your last downloaded version.
A button does nothing
Describe the symptom, not the fix: what you clicked, what you expected, what happened. Then ask Claude to fix only that.
I've hit my usage limit
Download what you have, click through it, and plan your next 2–3 prompts. When your allowance refills, start a fresh chat with the latest HTML attached and a short summary, so you don't resend the whole history.
Pro has a 5-hour session limit and a weekly cap. Download what you have and click through it. Next time, start a fresh chat in your Project and attach the latest HTML, rather than continuing a long chat.
The response stopped half-way
Reply "Continue from exactly where you stopped." For big prototypes, ask for fewer screens and build one flow at a time.
It doesn't look like the design system
Ask Claude to replace any custom-styled parts with the design system's own components, then compare against the system's component pages. You're the quality check.