---
id: prj-write-a-build-readme-before-you-vibe-code
schema: project/1
layer: garden
title: Write a one-page build plan before you ask AI to make an app
outcome: A one-page plan for a small app, with what it must do, what it leaves out and three milestones, that you can hand to any AI app builder so the first version matches what you meant and is easy to check.
roles: [operations staff, programme coordinator, small business owner, team lead, anyone who wants a small internal tool]
time_box: 25 minutes
tool_access: Any AI chat assistant (free tier is fine). Building the app afterwards needs an AI app builder or coding agent, but this card stops at the plan.
practises: [cmp-frame-the-outcome, cmp-describe-the-task, cmp-check-the-result]
concepts: [con-vibe-coding, con-smallest-useful-version, con-clarifying-questions, con-verification]
skills: [frame-the-problem, build-brief, prove-it]
builds_on: [prj-plan-with-clarifying-questions]
resources: [res-other-peter-yang-12-rules-vibe-coding, res-other-peter-yang-20-claude-code-tips]
source:
  url: https://creatoreconomy.so/p/12-rules-to-vibe-code-without-frustration
  claim: Asking the AI to write a short readme with requirements and a few milestones, and to plan before it codes, makes building apps with AI less frustrating for people who don't code.
  evidence: Peter Yang's account of more than 50 hours building apps with AI as a non-engineer. It is a practitioner's tips list, not a study.
  limits: The source shows his rules, not this exercise. The sample app, prompts, checks and failure test below are this card's own. Whether a written plan saves time on your build has not been measured.
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: Turns the first of his rules (plan before code) into a single-sitting exercise with a check and a failure test, without needing to install anything.
---

## Why this matters

When you ask an AI to "make me an app", it fills every gap with its own guesses, and you find out what it guessed only after it has built the wrong thing. Fixing the plan is quicker than fixing the app. A one-page plan also gives you a list to test the app against.

## Smallest useful build

A one-page plan with five parts: who uses the app, the job it does, three things it must do, two things it will not do yet, and three milestones, each with one test you can run yourself. About 25 minutes in any AI chat.

## Sample inputs

An invented request from a made-up tenants' association:

```text
We lend out tools (ladders, drills, a pressure washer) to about 60 households. Right now it's a paper sheet on the noticeboard and things go missing. I want a simple app where people can see what's available, book a tool for a day or two, and I can see who has what.
```

If you use your own idea, keep it to one small job and leave out names and real addresses.

## Steps

1. **Write your one line first.** Before you open the AI tool, finish this sentence yourself: "The app is working if ___." For the sample, something like "I can see in ten seconds who has the pressure washer."

2. **Ask the AI to question you, not build.**

   ```text
   I want to build a small app with an AI app builder, but not yet. First help me write a one-page plan. Here is my idea: [paste idea]. Ask me up to six questions, one at a time, about who uses it, what it must do and what can wait. Don't suggest features I haven't asked for. Don't write any code.
   ```

3. **Turn the answers into the plan.**

   ```text
   Now write the plan on one page with these headings: Who uses it, The job it does, Must do (three items), Not yet (two items), Milestones (three, smallest first, each with one test I can run myself without reading code). Use only what I told you. Mark anything you assumed with [assumed].
   ```

4. **Cut it down.** Read the plan against your one line from step 1. Remove anything that doesn't serve it, settle every [assumed], and make sure milestone 1 could be built and tested in an afternoon.

## Check it

- Your step 1 sentence is something milestone 1 or 2 lets you test.
- "Not yet" has at least two items, and the sample plan leaves out things like payments, logins for every household or reminders by text.
- Every milestone test is something you can do by clicking, such as "book the drill for Saturday and see it marked as out", not "the database works".
- No [assumed] lines are left.

## Break it

Go back to the same chat and add: "Oh, and it should handle deposits and late fees." Does the AI fold this into the must-do list and grow milestone 1, or does it ask whether this belongs in "Not yet"? If it grows the plan without asking, write a one-line rule you would give it next time, such as "Ask before adding anything to Must do."

## Reflect

- Which question from the AI changed your plan the most?
- What did you move to "Not yet" that you were tempted to keep?
- If someone else built from this plan, what would they still have to guess?

## Next iteration

Give the plan to an AI app builder and ask it to build milestone 1 only. Run your milestone test yourself before you ask for anything else. Keep the plan file and update it as you go, so each new chat starts from the same page instead of from memory.