Vibe Coding Prompts: 30 to Copy for Your First App in 2026
Vibe coding prompts you can copy: 30 messages for planning, building, fixing, securing and launching your first app, based on what 8 AI tools officially advise.

The best vibe coding prompts do three things: ask the AI to plan before it builds, change one thing at a time, and say what must stay untouched. Lovable, Bolt, v0, Replit, Claude Code and OpenAI's Codex all give that advice in their official guides. This page turns it into 30 prompts you can copy, in the order a first app needs them: plan, build, data, payments, fixes, security and launch.
You typed "build me a booking app for my salon" and got something that looked finished in two minutes. Then you asked for one small change and the calendar broke. You asked for a fix, and the fix broke the login. Forty messages later you are out of free credits and further from a working app than you were at message three. Nothing is wrong with you or the tool. The first message was a wish, not a brief. Good vibe coding prompts are short briefs, and the people who make these tools have published what a good one looks like.
Below: how we checked this guide, what 8 tools advise and charge, six rules they agree on, the 30 prompts by stage, three real failures a prompt would have prevented, and where an agent saves you the prompting altogether.
Prefer to describe the job once? On Jobbit Plan mode shows the steps before anything is built, and the agent hosts the finished app. Free plan, 100 credits a day, no card.
How we checked this guide
- Official guides, read on 14 November 2026. The prompting pages published by Lovable, Bolt, v0, Replit, Cursor, Anthropic for Claude Code, OpenAI for Codex and Google for AI Studio.
- Prices from each vendor's own pricing page on the same day. Most show a dollar sign without naming the currency, so we print them as shown.
- Every prompt is ours, written to follow that advice. Where a vendor gives a short example of its own, we say so.
- Failures come from named sources: Fortune, The Register, the security firm Wiz and the vendors' own documentation.
- Tested for sense, not for magic. A prompt cannot guarantee a result. It raises the odds and makes mistakes easier to undo.
What 8 vibe coding tools advise and charge
| Tool | Free allowance | Cheapest paid plan | What its official guide stresses |
|---|---|---|---|
| Lovable | 5 build credits a day, 30 a month | Pro from $25 a month | Let it question you first, one change at a time, paste the exact error |
| Bolt | 1 million tokens a month | Pro $25 a month | Describe the job, name the page, say what must not change |
| v0 | 7 messages a day | $20 or $30 a month, its two pages differ | Draft a requirements document first, build in stages |
| Replit | Starter plan with a daily allowance | Core $20 a month | Plan first, save a checkpoint after each working step |
| Claude Code | Not on the free plan | Claude Pro $20 a month | Say what "fixed" looks like, ask for evidence |
| OpenAI Codex | Limited use on ChatGPT Free and Go | ChatGPT Plus | Give the steps to reproduce a bug and how to verify the fix |
| Google AI Studio | Free tier | Pay per use | Start in plain language, keep keys on the server |
| Cursor | Free with limits | Plans under revision, check its page | Checkpoints before risky changes; built for developers |
Two notes. Cursor and Claude Code are aimed at people comfortable with code files, so they are a poor first stop. And Google's Firebase Studio no longer accepts new sign-ups, according to its own documentation, so skip guides that recommend it. For a fuller comparison see the best vibe coding tools.
Six rules the tool makers agree on
- Plan before you build. v0 suggests having the AI draft a requirements document. Anthropic suggests asking it to interview you until it can write a specification. Lovable and Replit both have a Plan mode for this.
- One change per message. Bolt and Lovable say it in almost the same words: make one change, check it works, then make the next.
- Say what must stay the same. The most common disaster is a fix that rewrites something that worked.
- Paste the real error. Every guide asks for the exact message or a screenshot, plus what you expected and what happened.
- Save before risk. Checkpoints, version history, a duplicate of the project: every tool has one. Use it before any big change.
- Stop clicking "fix". Lovable advises trying the automatic fix once or twice, then stopping to find the root cause. Bolt says the same about its "Attempt fix" button.
If the idea is new to you, start with what vibe coding is. Collins Dictionary made it the word of the year for 2025, defining it as "the use of artificial intelligence prompted by natural language to assist with the writing of computer code".
Prompts 1 to 5: plan before a line is written
Replace the words in square brackets with your own.
- The interview. "I want to build [a booking app for a two-chair hair salon]. Before you build anything, ask me questions one at a time until you understand who uses it, what they do first and what I need to see as the owner. Then stop."
- The one-page brief. "Write a one-page plan for this app in plain English: who it is for, the 3 things a user can do, the pages, and what is deliberately left out of version one. Do not write any code yet."
- The smallest version. "From that plan, list the smallest version that would be useful to one real customer this week. Move everything else to a later list."
- The data in plain words. "List the information this app will store, in plain English, as tables with their columns. For each table, say who is allowed to see each row: the customer, the owner, or anyone."
- The look. "Use this style for every screen: [calm, light background, one accent colour, large buttons, no stock photos]. Save this as a standing instruction so I do not have to repeat it."
Prompts 6 to 11: build one piece at a time
- The first screen. "Build only the [booking page]: a list of services with prices, a date picker showing free slots, and a confirm button. Use realistic example text, not placeholder words. Build nothing else yet."
- The next piece. "The booking page works. Now add [the owner's day view] as a separate page. Do not change the booking page."
- The small change. "On the [booking page], change [the confirm button text to 'Book and pay later']. Change nothing else on this page or any other."
- The phone check. "Make the [booking page] work on a phone: one column, buttons at least a thumb wide, no sideways scrolling. Do not change how it looks on a computer."
- The awkward cases. "For the [booking form], handle these cases with a clear message: no slots left, a date in the past, an email typed wrongly, the connection dropping while saving."
- The explanation. "Explain what you just changed in plain English, in five bullet points, and tell me which parts of the app it could affect."
Prompts 12 to 15: accounts and data
- Sign-in. "Add user accounts with email sign-in. Customers see only their own bookings. The owner account sees all bookings. Tell me how you enforced that."
- The lock on every table. "For every table in the database, switch on row-level security and write a rule for each action: read, add, change, delete. Then show me a table of who can do what."
- The stranger test. "Act as a visitor who is not signed in. List every page and every piece of data you can reach. Then do the same as a signed-in customer trying to see another customer's bookings."
- Test data out. "List all the example data you created. Remove it, and confirm that the app starts empty for a new owner."
Prompts 16 to 18: taking payments
Lovable's Stripe guide is firm on two points: enter the payment key in the tool's secure connection form, never in the chat, and try everything in Stripe's test mode first, because a live key in a preview means real charges.
- A single payment. "Add a one-time payment of [£20 deposit] at the end of the booking flow using Stripe in test mode. I will enter the key in the secure settings, not here. Show a clear success page and a clear failure page."
- The proof. "Walk me through testing this payment with Stripe's test card, step by step, and tell me where in the Stripe dashboard I will see the result."
- Going live. "List every step needed to switch from test payments to live payments, in order, and what I must check after each one. Do not switch anything yet."
Prompts 19 to 23: when something breaks
- The bug report. "On the [booking page], when I [choose a Saturday], I expected [the free slots] but got [an empty list]. Here is the exact error: [paste it]. Find the cause before changing anything and tell me what it is."
- The root cause. "You have tried to fix this twice. Stop. List three possible causes, say which is most likely and why, and how we can check each without changing the code."
- The safe fix. "Fix only that cause. Do not tidy, rename or improve anything else. Then tell me exactly which files changed."
- The proof it works. "Show me the evidence that it is fixed: what you tested, what happened, and one case you checked that should still fail."
- The fresh start. "Summarise this project so far: what is built, what is broken, what we decided. I will paste this into a new chat." Bolt recommends this when a long conversation starts to drift.
Prompts 24 to 27: security before anyone else sees it
- The review. "Review this app's security as if you were trying to break in. Check the sign-in flow, every route that reads or writes data, and file uploads. List problems by severity." Lovable's own prompt library starts with a shorter version: "Review my app's security."
- The keys. "Search the whole project for passwords, keys and tokens. Tell me where each one is. None may be in code that runs in the browser. Move any you find to the secure secrets store."
- The personal data list. "List every piece of personal information this app collects, where it is stored and who can see it. Tell me what a customer would need in a privacy notice."
- Two separate worlds. "Confirm that the version I test with and the live version use separate databases. If they do not, tell me how to separate them before we go further."
Our guide to vibe coding mistakes and security covers these risks in more depth. An app holding sensitive data needs a review by a professional, which Lovable's documentation says too.
Prompts 28 to 30: launch and look after it
- The launch list. "Give me a launch checklist for this app: my own domain, secure connection, a backup, a contact address, a privacy page, and what it will cost each month to keep running."
- The handover note. "Write a one-page note for someone who has never seen this app: where it is hosted, how to sign in as the owner, how to export the data, and who to call if it stops."
- The monthly check. "List five things I should check on the first of each month to know this app is healthy, and how to check each without technical knowledge."
Putting the result on your own address is its own job. See how to put an AI-built app on your own domain.
Three failures a prompt would have prevented
- The deleted database. In July 2025 Fortune reported that an AI agent on Replit wiped a company's live database during a code freeze. The chief executive called it "unacceptable" and Replit added separate test and live databases. Prompt 27 asks for that separation on day one.
- The door left open. In July 2025 the security firm Wiz showed that private apps on one builder could be entered by anyone who knew the app's public identifier. The vendor fixed it within a day. Prompts 13 and 14 make you look before a stranger does.
- The wiped drive. In December 2025 The Register reported that a photographer asked an AI coding tool to clear a project cache and lost the contents of a whole drive, while running in a mode that executes commands without asking. Rule 5 applies: save first, and leave auto-run switched off.
How an AI agent saves you the prompting
Assistants such as ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse will all write code from a description. The 30 prompts above exist because someone still has to manage the build. An agent built for finished work takes over more of that management.
- It plans first by design. Plan mode on Jobbit lists the steps and waits for your yes.
- It builds and hosts in one place. There is no separate hosting account to set up. Go, at £7 a month, keeps one small app live on your own domain. See Jobbit's plans.
- You can watch. The agent's browser is visible as it works and you can take control.
- It keeps the instructions. Projects hold your standing rules and files, so prompt 5 is typed once.
- A person can finish it. When the job needs a professional, a security review for instance, Jobbit's human network takes it with payment held in escrow.
If you are still choosing an approach, compare vibe coding with no-code and low-code.
Frequently asked questions
What is a good prompt for vibe coding?
A good prompt names one job, the place it happens and what must not change. "On the booking page, change the button text to Book now, and change nothing else" beats "improve the booking page". For anything bigger, ask the AI to plan first and approve the plan before it builds.
How do I start a vibe coding project?
Start with questions, not code. Ask the AI to interview you about who uses the app and what they do first, then to write a one-page plan with a deliberately small first version. Prompts 1 to 3 on this page do exactly that.
Why does the AI keep breaking things that worked?
Because the request did not say what to leave alone, and because long conversations drift. Make one change per message, name the page, add "change nothing else", and save a checkpoint after each working step. If the chat has grown long, ask for a summary and start a fresh one.
Do vibe coding prompts work in every tool?
The ideas do, because the official guides from Lovable, Bolt, v0, Replit, Anthropic and OpenAI agree on them. The buttons differ: Plan mode, checkpoints and secure key storage have different names in each tool.
Is an app built with prompts safe to launch?
Only after you have checked it. Ask for a security review, switch on row-level security for every table, keep keys out of browser code and test with an account that should not have access. If the app handles payments or sensitive personal data, pay a professional to review it.
Can I build an app without writing prompts like these?
You can hand more of the management to an agent. On Jobbit you describe the result, approve the plan and the agent builds and hosts it, with a human professional available when needed. Try it free with 100 credits a day and no card.