How to Prompt Cursor So It Doesn't Wreck Your Code
Cursor agents fail because your prompt was a product pitch. One task, a file list, a done check — plus when to kill the chat and start a fresh thread.
Who This Is For
You opened Cursor, pasted the entire startup idea, and said "build it."
Forty-five minutes later you have auth stubs, a settings page, three half-wired APIs, and a git diff you are afraid to read.
The model is not drunk. You handed it a company. Companies do not fit in one context window. Weekend products fit in one magic moment — see the one-feature product.
This article is about the message you type in the chat. The standing instruction file is a different layer: AGENTS.md. Put durable facts there. Put this task in the prompt.
If you do not have a repo yet, pick a scoped brief from startup ideas so the first prompt has a boundary.
The Lie: Longer Prompts Are Better
A product pitch feels thorough. To the agent it is permission to invent surface area.
Cursor's own 2026 agent-practice wave (and every serious AGENTS.md guide) rhymes on the same point: static context should be short. Augment's 2026 AGENTS.md write-up argues for ~150–200 lines of standing rules, not a novel. RockB's Cursor agent notes say keep AGENTS.md under ~500 words because it loads every run. Your chat prompt should be even meaner: one outcome, not a manifesto.
If you already dumped 2,000 words of "vision" into .cursor/rules, the chat cannot save you. Cut the standing files first.
The Prompt Shape That Works
Copy this skeleton. Fill the brackets in plain English. Do not add a fifth section "also make it beautiful."
Task: one sentence. The user-visible change.
Touch: files or folders allowed. "Only app/(app)/onboarding/ and tests for that route."
Do not: two or three hard nos. "Do not add a new table. Do not install a component library. Do not refactor auth."
Done when: a command or a click path. "npm test passes. I can submit the form and see the toast."
Verify: "Run the test. If it fails, fix it. Paste the command output."
That last line is how you stop the agent from claiming victory on a type error.
Bad Prompt vs Good Prompt
Bad
Build a complete SaaS for piano teachers with subscriptions, dashboards, AI lesson plans, parent emails, and a mobile-ready design system. Use best practices.
That is how you get a settings page.
Good
Add a page at /plans/new with a textarea and a Generate button. On submit, call the existing
generatePlanConvex mutation and render the three returned fields. Do not add auth, billing, or email. Done when I can paste text, click Generate, and see the plan without a full-page reload. Run the relevant tests.
Same weekend. Opposite blast radius.
Five Rules for the Chat Itself
1. One job per thread
When the task changes — "now add Stripe" — new chat. Saturated threads start "fixing" files you did not mention. Cursor's 2026 best-practice videos are blunt: start a fresh agent when context gets mushy.
2. Name the files
"Look around and do the right thing" is how you get a second Button.tsx. If you do not know the files, your first message is: "Where is signup handled? List paths. Do not edit."
3. Plan when the change spans more than two folders
Switch to plan/ask, get a file list, then implement. Multi-folder improvisation is how schema and UI drift apart.
4. Tests are the spec
If you cannot write a test, you cannot describe done. For UI, a Playwright click path or a "I will click X and see Y" sentence is the spec. Agents that cannot check their work will cheerfully regress you.
5. You still read the diff
Non-negotiable. Especially auth, payments, and anything that talks to the network. Vibe coding security is what happens when you skip this.
What Belongs in AGENTS.md vs the Prompt
| Lives in AGENTS.md / rules | Lives in this chat |
|---|---|
npm run test command | "Add the empty-state on /plans" |
| "We use Convex, not Prisma" | "Do not touch schema.ts" |
Never commit .env | Today's definition of done |
Do not paste the idea thesis into both. The thesis belongs in an issue or a brief. The agent needs the diff.
A Weekend Cadence
Friday night: write the magic-moment sentence. Open the idea page if you used startup ideas. One prompt: scaffold the three screens, nothing else.
Saturday morning: prompt #2 is the core mutation/API only.
Saturday afternoon: prompt #3 is the UI wired to that API. No dashboard.
Sunday: prompt #4 is deploy + the one email. New chats for each.
Four prompts. Not one "build my startup" novella.
When It Starts Wrecking Things
Symptoms:
- Files you did not name in the diff
- New dependencies you did not ask for
- Reformatting the whole repo
- "While I was here I also..."
Stop. git checkout the junk. New chat. Smaller prompt. If it keeps "helping," your standing rules are bloated or contradictory — trim those before you type another word.
FAQ
Should I paste stack traces raw?
Yes, plus the command you ran. "It doesn't work" wastes a turn. The error is the spec.
Is Composer/agent mode always better than inline edit?
For multi-file, agent. For a 10-line fix in a file you have open, inline. Matching tool to blast radius is the whole game.
Can I reuse a mega-prompt as a Cursor command?
Only after it has succeeded three times. Then it is a command. Until then it is superstition.
TL;DR
Standing rules stay short. Each chat is one task, named files, explicit nos, and a checkable done. Read the diff.
Want a product small enough that the first prompt fits on a card? Start at startup ideas and build the how-it-works list, not a platform.