Last week I wrote about the workflow: stop trying to write the perfect plan yourself, and let the AI build it with you through a conversation. A few people asked the obvious follow-up — okay, but what do I actually type?
Fair question. A workflow is useless if you don't know the words that trigger it. So here are the exact prompts I use. They work in Claude Code's Plan Mode, but nothing here is tool-specific — paste them into any agent before it writes a line of code.
The problem with "just describe your idea"
If you open a chat and say "build me an app that does X," the agent does what agents do: it fills every gap with an assumption and starts coding. You get a plan you never agreed to.
The fix is to change the agent's job before it starts. Don't ask it to build. Ask it to interview you, propose a structure, and argue with itself. Here's how.
The master prompt
This is the one I paste first, on an empty project, before anything else:
I have an idea for an app and I want to plan it with you before any code is written.
Do NOT write code yet. Your job right now is to help me produce a short, clear plan.
Here is the rough idea:
[describe your idea in 2-4 sentences — what it does and who it's for]
Before you propose anything, ask me the questions you need to remove your own
assumptions: about the users, the core workflow, the data, and the screens.
Ask them in small batches so I can actually answer.
Once you have enough, propose a short plan with:
- the main user and their main job
- 3-5 core flows
- the key data objects
- the main screens
- the unhappy paths that could corrupt data, leak permissions, or cost money
- a "not building yet" list of things we are deliberately leaving out
Keep it short and rough. We will refine it together.
The important parts: "do not write code yet" stops the rush. "ask me the questions you need to remove your own assumptions" turns the agent from a guesser into an interviewer. And the "not building yet" list is what keeps it from inventing scope later.
The smaller prompts for refining
Once it gives you a first draft, you steer with short follow-ups. These are the four I reach for most.
Make it push back instead of agreeing:
For each major decision in this plan, give me the pros, the cons, and at least
one alternative. If something is a bad idea, say so and tell me why. Do not just
agree with me.
Pressure-test a single choice:
You suggested [X]. Why is that the right call here? What would you lose by doing
the simplest possible version instead? Give me the trade-off, not reassurance.
Lock in the boundaries:
Turn the "not building yet" list into clear non-goals. Phrase each one as a rule
I can drop into my project instructions so you don't reintroduce it later.
Confirm you're done planning:
Are there any questions left that, if I answered them wrong, would send us down
the wrong path? If yes, ask them now. If no, summarize the final plan in one
short block so I can approve it.
That last one is how I know it's time to stop planning and start building — instead of guessing, I let the agent tell me whether anything important is still unanswered.
Why this works
None of these prompts are clever. They just reassign the agent's role at each stage: interviewer first, critic second, scope-keeper third, and only then a builder. The frustration in my old workflow came from skipping straight to "builder" and letting it improvise everything before it.
Try this
Next time you start something new, don't describe the app. Paste the master prompt, answer the questions honestly, and force it to argue with you before you approve anything.
You're not writing the plan. You're running the conversation that produces it — and that's a much easier job.
Top comments (0)