Code · The What2Prompt engine

Vibe Coding Prompts for Lovable, Cursor & Claude Code

Vibe coding starts with describing what people should be able to do, then checking that the app actually does it. These prompts for Lovable, Cursor and Claude Code break a build into small milestones, evidence-led bug fixes and “don’t break existing code” boundaries. Use the tool that can inspect your project, and keep permissions, tests and human review in the loop.

See all prompt guides →
2Which AI will you paste it into?
Short and punchy.
1 credit per prompt

Works with ChatGPT, Claude, Gemini, Midjourney, Suno and more.

Copy · Paste · Done
STEP 3 · YOUR PROMPT
Prompt for Claude CodeQuickExample

Role

Act as a coding agent inspecting the existing project before implementation.

Context

I want a task tracker for a small team. A signed-in member should create a task, see only permitted tasks and mark one complete. Use the current framework, design controls and access rules. I will supply the real project context; do not assume file paths.

Task

1. Locate the existing sign-in, task data and UI patterns.
2. Propose the smallest milestone covering create, read and complete.
3. Identify missing business rules before editing.
4. Implement only the approved milestone, with empty, loading and failed-save states.

Format

Return a focused plan, changed files and concrete test results for creation, completion and a different member’s access. Separate actual checks from unverified ones.

Guardrails

No unrelated redesign, new stack, invented integrations or broad permission changes. Do not put secrets in source. Ask before destructive operations or changing payment behavior.
5 free prompts

Example prompts

Example

Describe an app as a first working milestone

Role

Act as a coding agent inspecting the existing project before implementation.

Context

I want a task tracker for a small team. A signed-in member should create a task, see only permitted tasks and mark one complete. Use the current framework, design controls and access rules. I will supply the real project context; do not assume file paths.

Task

1. Locate the existing sign-in, task data and UI patterns.
2. Propose the smallest milestone covering create, read and complete.
3. Identify missing business rules before editing.
4. Implement only the approved milestone, with empty, loading and failed-save states.

Format

Return a focused plan, changed files and concrete test results for creation, completion and a different member’s access. Separate actual checks from unverified ones.

Guardrails

No unrelated redesign, new stack, invented integrations or broad permission changes. Do not put secrets in source. Ask before destructive operations or changing payment behavior.
Example

Fix a bug without regenerating the app

Role

Act as a debugging agent working from current source and observed behavior.

Context

The saved task list shows oldest items first after refresh. Expected: newest creation time first, unchanged owner access. I will supply a reproduction, two sample timestamps and the actual error if there is one.

Task

1. Reproduce the ordering issue and locate the query and render path.
2. Explain the cause from the code.
3. Apply the smallest change preserving existing pagination and permissions.
4. Add a regression test proving the newer timestamp comes first after reload.

Format

Report the cause, scoped diff, test input and observed result. List any browser or data checks you could not run.

Guardrails

Do not rewrite the data layer, alter access rules or claim a test passed without running it. Use sanitized fixtures instead of private customer records.
Example

Add a feature without breaking working code

Role

Act as a coding agent extending the established project patterns.

Context

The task tracker already creates, lists and completes tasks correctly. Add a favorite toggle only. Preserve those flows, existing public URLs, data ownership and the current visual style.

Task

1. Inspect the task types, update flow and existing tests.
2. Plan a compatible favorite state and save behavior.
3. Implement the toggle using the existing controls.
4. Test adding and removing a favorite, failed saves and an unauthorized record.
5. Run regression checks for create and complete.

Format

Return the plan, changed files, successful checks and remaining risks. If a schema change is needed, explain it before applying it.

Guardrails

No parallel architecture, broad refactor, dependency replacement or destructive data migration. Protect secrets and never choose another user’s account from client input.

Describe the app through a user workflow

An app brief should begin with the person using it and one action they need to complete. “A task tracker for a small team” is a category; “a member creates a task, sees it after reload and marks it complete” is a workflow. Add what happens when input is missing, saving fails or the member lacks permission. Those details guide both interface behavior and the data model. Keep the first version small enough that you can test it end to end.

Give only business facts you actually know. If pricing, roles or retention rules are undecided, mark them as questions rather than allowing the coding agent to invent defaults. Name the existing framework and current patterns when a project already exists. Let the tool locate files it can inspect; do not guess paths from another codebase. A screenshot can communicate appearance, but it cannot establish permissions, persistence or what a payment is supposed to do.

Build step by step instead of accumulating screens

A vertical slice includes the whole action: input, validation, save, feedback and read-back. Complete that slice before adding a dashboard full of placeholder controls. For a task tracker, creating one task with owner-scoped reads is a better first milestone than five decorative pages. Ask for the empty state, an in-progress state and a failure message as part of the same feature. This makes the workflow usable and exposes missing behavior early.

After the slice works, name the next small extension and the checks it must preserve. Favorite state should not break completion; a new filter should not broaden owner access; pagination should not silently change ordering. Require actual inputs and expected results. The agent should report which checks ran and which could not. Do not treat a compilation result as proof of a multi-step user flow, and do not add more features to avoid investigating a failed core action.

Debug from evidence, not a more forceful request

A useful bug report has a reproduction, an expected outcome and an observed outcome. Include the exact error text and a small sanitized example. “After refresh, the newer task appears below the older one” directs an investigation; “the dashboard is broken” invites speculation. Ask the agent to locate sibling paths that share the same assumption, such as filters or sort logic, before applying a repair. Fixing only the visible symptom can leave the same defect elsewhere.

Keep the repair scoped to the cause. A query-ordering defect should not trigger a new data library or a redesign. Request a regression test with two concrete timestamps and the expected sequence. If the issue depends on login or persistence, reproduce it in that state rather than checking only an empty screen. Never paste credentials, raw customer records or private access tokens into the debugging conversation. Use restricted test accounts and disposable data where possible.

Make “don’t break existing code” testable

List the behaviors that must remain true: public URLs, owner boundaries, saved data shape, payment rules and existing create or update flows. Ask the coding tool to read the current implementation and reuse the established controls and modules. When a structural change is necessary, have it explain compatibility and migration risks before proceeding. A guardrail should protect a real invariant, not simply discourage every change that might improve maintainability.

Prompts are instructions, not enforcement. Keep destructive operations subject to confirmation, restrict credentials and review changes involving identity, payments or private data. Run focused tests plus checks for the preserved flows. Inspect the resulting app at the actual viewport and with a signed-in user when the workflow requires one. A careful agent should admit missing evidence rather than convert a successful code edit into an unsupported claim that everything is production-ready.

What makes a good vibe coding prompt brief

  1. Describe the user, action and observable result, not only the app’s name or an aesthetic reference.

  2. Build one complete vertical slice before adding several unfinished screens or integrations.

  3. Tell the agent to inspect existing code and reuse established data, access and design patterns.

  4. Include failing cases, owner boundaries and concrete acceptance tests alongside the happy path.

  5. Keep destructive actions, credentials and payment changes under explicit review; prompt guardrails do not enforce permissions.

Bad prompt vs What2Prompt prompt

Bad prompt

“Build my whole app and make it perfect”

What2Prompt prompt

Role

Act as a coding agent inspecting the current repository.

Context

An existing task tracker; add only a favorite toggle while preserving create, complete and owner access.

Task

Locate the current update pattern, implement a compatible toggle and cover add, remove and failed-save behavior.

Format

Plan, scoped changes and actual regression results.

Guardrails

Do not replace the framework or data layer, broaden permissions, expose secrets or claim unrun tests passed.

The weak request offers no milestone or definition of done. The stronger prompt names a bounded change and the working flows to preserve. Lovable prompts, Cursor prompts and Claude Code prompts can share that structure, but the available project access and execution environment differ. Verify the result in the actual tool rather than assuming the wording guarantees safe code.

Before accepting a build, use the app as its intended user. Create a real sample record, refresh, change it and confirm it reads back correctly. Test a failed action and, when permissions matter, a different account. Read the diff for accidental stack replacements or removed checks. Keep a list of unverified flows; a generated completion report is not proof of deployment, billing readiness or private-data isolation.

Reference: Claude Code best practices ↗

Questions about vibe coding prompt

What is a useful vibe coding prompt?

It explains who uses the app, what action they take, what should happen and how success will be checked. Include constraints the repository cannot reveal, such as a business rule or required public URL. Start with a small working flow, then extend it after verification. A visual description alone does not define data or access behavior.

Can I reuse prompts across Lovable, Cursor and Claude Code?

Reuse goals, constraints and acceptance cases, but adapt to the tool’s actual project access and workflow. Do not ask a tool to inspect files it cannot see or run commands it cannot execute. At the time of writing, November 2026, capabilities and plans vary; check the selected environment before depending on a feature.

How do I stop AI from breaking existing code?

Explicitly preserve named working flows, require inspection before edits and ask for a focused diff plus regression checks. Keep backups and restricted permissions, and review risky changes yourself. Instructions help direct the agent, but only actual tests and enforced access rules provide evidence and boundaries. “Don’t break anything” alone is not a test plan.

What2Prompt prompt engine, with the W2P mark on cobalt.

Give your next idea a better brief.