Skip to content

Vibe coding prompts

By Mari Luukkainen · 9 min read · Last updated

On this page

These 32 vibe coding prompts cover the tasks you repeat most: starting a project, adding a feature, fixing a bug, writing tests, publishing and checking security. Replace the text in square brackets, paste the prompt into your tool and read the result before you accept it.

Each prompt is built on three habits: state the outcome and the check, set a boundary on what may change, and ask for the output you can verify. The prompts work in browser builders, editors and terminal agents. Keep the ones that work for you in a text file. For a full worked example, see how to vibe code.

Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok

Start a project

A first prompt sets the shape of everything after it. State who uses the app, what each screen does, what is out of scope and ask for a plan before code.

1. New app with a plan first
Create a web app called [name].
Who uses it: [one sentence].
What it does: [one screen at a time].
Out of scope: accounts, payments, a database (change this if needed).
Use Next.js with TypeScript and Tailwind CSS. Mobile first.
Before you write code, list the files you plan to create and wait for my OK.
2. Turn a vague idea into a spec
I want to build [vague idea]. Ask me up to eight questions that you need answered before building it. Ask them one at a time, wait for each answer, then write a one-page spec with: users, screens, data that is saved, and what is out of scope.
3. Goal prompt (outcome and definition of done)
Goal: [outcome] for [who].
Done means: [two or three checks you can verify, with numbers].
Why it matters: [what it feeds].
Decide yourself: [layout, names, small choices].
Ask me about: [database changes, anything that costs money].
Scope first: read the project, tell me the shape you propose, then build.
4. Copy the style of an existing page
Look at app/about/page.tsx and the components it uses. Build app/team/page.tsx in the same style: same spacing, same heading sizes, same components. Do not create new styles.
5. Start from a design brief
Design brief: warm off-white background, near-black text, one accent colour (#b93a07), a serif for headings, a sans-serif for body text at 18px. No gradients, no shadows, no icons in rounded squares, no centered hero. Left-aligned. Apply it to the home page only and show me before touching other pages.
6. Audit a page for the generic AI look
Look at [page]. List every element that looks like the default output of an AI tool: gradient hero, three identical feature cards, icons in rounded squares, centred everything, purple accents, emoji as bullets, filler headlines. For each one, propose a plainer replacement that fits this brand: [two adjectives, one colour]. Use only the real text I give you and do not invent testimonials, numbers or logos. Show the list before you change anything.
Your first Claude Code prompt · Claude

Add a feature

Keep each feature to one prompt with a boundary, so you can tell what changed.

7. Add a feature with a boundary
Add a [feature] to [page]. Behaviour: [what the user does and what happens]. Do not change anything else on the page and do not add dependencies. List the files you changed.
8. Add a form that handles bad input
Add a contact form with name (optional), email and message (at least 10 characters). Validate on the server, not only in the browser. Show the error under the field that is wrong. Add a hidden field that bots fill, and reject submissions that fill it. Limit to five submissions per hour per IP address.
9. Add search or filter
Add a text box above the list that filters items by title as I type. With no match, show "Nothing found". Keep the filter text in the URL so I can share the page. Add a test for the filter function.
10. Add a settings page
Add /settings with display name and email fields. Load the current values, validate on the server, show "Saved" after a successful save, and keep unsaved changes if validation fails. Link to it from the main navigation.
11. Take payments with Stripe and handle the webhook
Add Stripe Checkout for the product [name] at [price]. The secret key and the webhook signing secret go in environment variables, never in the code. Create the webhook route: verify the Stripe signature against the raw request body, reject anything unsigned, and store each event id so the same event is never processed twice. Mark an order as paid only from the webhook, never from the page the customer returns to. Tell me how to test it with the Stripe CLI.

Fix a bug

A bug prompt has three parts: what you did, what you saw, what you expected. Paste the exact error text.

12. Standard bug report
Bug: [what you did].
I saw: [exact message or behaviour].
I expected: [what should happen].
Reproduce it with a failing test first, then fix it. Only change what the bug needs. Tell me which file and line were wrong.
13. Paste an error from the terminal
This error appears when I run npm run build:

[paste the full output]

Explain in two sentences what it means, then fix it without disabling any checks or adding ts-ignore comments.
14. The AI keeps breaking the same thing
You have broken [feature] three times now. Before you change anything, read [file] and write down in three bullet points the rules this feature must follow. Add those rules as comments at the top of the file. Then fix the current problem without breaking any of them.
15. Works on my computer, fails when published
The published site shows [problem] but the local version works. List every difference between local and production that could cause it: environment variables, build output, file names with different capital letters, paths, server and browser code. Check each against the project and tell me the most likely cause before changing anything.
16. Find the cause, do not guess
Do not fix this yet. Add temporary logging to find where [value] becomes wrong. Run the app, show me the log output, tell me the line responsible, then remove the logging.

Write tests

Tests are the checks the tool can run on its own work.

17. Tests for existing code
Write tests for [file or function]. Cover: a normal case, an empty input, a very long input, and a wrong type. Run them. If any fail, tell me whether the test or the code is wrong. Do not change the code yet.
18. Test first, then fix
Write a test that fails because of this bug: [description]. Run it and show me the failure. Then fix the code and run the test again.
19. Check the user path
Write an end-to-end check for this path: open the home page, click Sign up, submit the form with a valid email, see the thank-you page. List what you used to run it and how I run it myself.

Understand and tidy the code

You do not need to read every line. You need to know what the parts are.

20. Explain a change
Explain the change you just made: which file, why, and what would break if I removed it. Name one thing in it that I should be able to explain to someone else.
21. Refactor without changing behaviour
Refactor [file] so that [problem]. Behaviour must not change. Run the existing tests before and after and show both outputs. If there are no tests for this code, write them first.

Publish, SEO and moving off a builder

Publishing fails in predictable places: the build, missing environment variables and wrong URLs.

22. Pre-publish check
Run npm run build and fix every error without turning checks off. Then list every environment variable the project reads, mark which are secret, and tell me which must be set at the host. Check that .env files are in .gitignore.
23. Metadata and sitemap
Add a title (under 60 characters) and a description (under 155 characters) to every page, an Open Graph image, a sitemap and a robots file. List each page with its title and description in a table.
24. My one-line SEO prompt
SEO the shit out of it.

Then list what you added in a table: titles, meta descriptions, Open Graph tags, alt texts, internal links. Remove anything outdated, such as a meta keywords tag. Do not choose keywords for me; I will give you the keyword for each page.
25. Move a project off a builder
I exported this project from [builder]. List everything that still depends on the builder's own services: hosted database, login, file storage, environment variables, build settings. For each one, propose a replacement and say what data has to be moved. Do not change any code yet.

Security and safety

Run these in a fresh session, so the tool does not defend its own work.

26. Secrets scan
Search the whole project and the Git history for API keys, tokens, passwords and connection strings. List every hit with file and line. Do not print the secret itself.
27. Permission check
List every page and API route. For each one write whether it returns user data and what check stops one user reading another user's data. Mark any route that has no check.
28. Input handling
Find every place where user input goes into a database query, a file path, a command or HTML. For each, say whether it is validated or escaped. Propose fixes but do not edit yet.
29. Rate limits
List every form and API route that sends email, creates records or checks passwords. For each, state whether there is a limit on repeated requests. Propose limits.
30. Row Level Security check
List every database table. For each one, say whether Row Level Security is enabled and list its policies in plain language. Mark any table without RLS, and any policy that lets one user read or write another user's rows. Then write a test with two users that proves user A cannot read user B's data. Do not change anything yet.

Instructions that persist

Rules you repeat in every prompt belong in the project's instruction file. Claude Code reads CLAUDE.md, or AGENTS.md when there is no CLAUDE.md. Codex, GitHub Copilot, OpenCode and Cursor read AGENTS.md. Cursor also has its own rules in .cursor/rules, and Copilot reads .github/copilot-instructions.md. The simplest setup is AGENTS.md as the source and a CLAUDE.md containing one line, @AGENTS.md.

31. Draft the instruction file
Read the project and write a short instruction file: the commands to run, test and build; the folder layout in five lines; and five rules based on how the code is written. Keep it under 40 lines. Mark guesses with a question mark.
32. Add a rule after a mistake
Add this rule to the instruction file in one line: [the mistake, written as a rule]. Keep the file under 40 lines by removing any line that is no longer true.

How to adapt these prompts

Replace the text in square brackets, keep the structure. If a prompt gives a result you do not like, change one part and run it again: tighten the scope, add a number to the definition of done, or add a rule about what not to touch.

Two rules from my own posts. From what I learned building prompt files: the model fills every gap you leave, so do not leave gaps, and define the output format. From the goals post: state the outcome, what done looks like in numbers and the reason, then hand over the decisions and mark the few that are yours. In the Herizon codebase I have 167 of these written down.

The SEO prompt above is from March 2025. I gave Cursor "SEO the shit out of it" on a minimal landing page and it produced titles, meta descriptions, Open Graph tags, alt texts and some internal link structure. It also added outdated meta keyword stuffing, and schema markup needed more prompting. I would not trust it to choose keywords, so I pick those myself (post).

Pick a tool from the tool overview, and see is vibe coding bad before you ship anything that holds other people's data.

Frequently asked questions

Do I need to start a prompt with a role such as senior developer?

Rarely. In my notes on prompt files I wrote that with newer models roles matter less than they used to. Keep one if you want a particular voice, and skip it if all you need is structured output.

Should I write please in a prompt?

It is not needed. Filler such as please or kindly adds length and no information. Spend the words on the outcome, the boundary and the check.

Can AI choose the SEO keywords for my site?

I would not trust it to. In my March 2025 test Cursor handled titles, descriptions and tags well, but choosing a keyword that can reach the top ten is still a human decision. Pick the keyword first, then prompt for the metadata.

Which file do I put repeated rules in for Cursor, Copilot and Claude Code?

Put them in AGENTS.md, which Cursor, Copilot, Codex, OpenCode and Claude Code can read. Cursor also reads .cursor/rules and Copilot reads .github/copilot-instructions.md. For Claude Code, add a CLAUDE.md containing the line @AGENTS.md.

Is it safe to paste my .env file into a prompt?

No. A pasted key now sits in a chat history and can end up in a commit. Paste variable names without values, and if a real secret was pasted, replace it with a new one at the provider.

Next guideIs vibe coding bad? →