Build a real app with AI without it falling apart (the phase method)

Everyone's AI apps break the same way: one giant prompt, forty minutes of generated code, then a collapse nobody can debug. To build an app with Claude Code that actually ships, you don't need better prompts. You need phases. This is the method I used to rebuild this entire site.

What you'll have when you're done

• A master plan file that splits your app into 5-7 phases with hard stop points
• A PROGRESS.md tracker the AI reads at the start of every session, so nothing gets forgotten between sessions
• A checkpoint ritual where you review and approve before the AI continues
• The exact standing rules that keep the AI from silently going off-plan
• [SCREENSHOT: a PROGRESS.md with phases checked off]

Before you start

• Claude Code installed. It's Claude running in a terminal (the text window developers use), and it can read and write real files on your computer. Included with the Claude Pro plan at $20/month.
• An app idea you can describe in a paragraph.
• 45 minutes for planning. No code gets written today, and that's the point.

Step 1: Write the master plan before any code

Create a file called MASTER_PLAN.md in your project folder. This is the one prompt that starts every phase. Paste this and fill in your app:

My app: [one paragraph: what it does, who it's for, what "done" looks
like]

We build in phases. One phase per session.

Rules:
- At the end of each phase, update PROGRESS.md and STOP for my review.
  Never continue into the next phase without my approval.
- Start every session by reading PROGRESS.md before doing anything.
- If anything in the code conflicts with this plan, flag it in
  PROGRESS.md. Don't silently pick one.
- Commit the work at every checkpoint.

The phases:
Phase 0: Audit. Read everything that exists, write down what's there
  and what's missing. No code changes.
Phase 1: Skeleton. The app runs end to end with placeholder content.
Phase 2: Design. Three visual directions as static comps. I pick one,
  then it gets locked.
Phase 3: Core build. The main feature working with real data.
Phase 4: Polish. Every page reviewed, mobile checked, errors handled.
Phase 5: Content and launch. Real content in, launch checklist done.

Adjust the phase list to your app, but keep phase 0 as a look-don't-touch audit and keep every phase small enough that you could review it in 20 minutes.

Step 2: Create the tracker

The tracker is the AI's memory between sessions. Create PROGRESS.md next to the plan:

# PROGRESS

Read this first, every session, before doing anything.

## Phase status
- Phase 0: not started
- Phase 1: not started
[one line per phase]

## Checkpoint log
[the AI appends an entry here at the end of every session: what was
done, what's incomplete, what needs my review]

## Conflicts
[anything found in the code that contradicts the plan gets flagged
here, never silently resolved]

This file is why the method works. A fresh session has no memory of last night, but it reads this file and knows exactly where the project stands. Mine currently holds the full history of this site's rebuild, checkpoint by checkpoint.

Step 3: Run phase 0, the audit

Open Claude Code in your project folder and send your first real prompt:

Read MASTER_PLAN.md and PROGRESS.md. Then run Phase 0: audit what
exists in this folder (or confirm it's empty), write down what we have
and what's missing for the plan, update PROGRESS.md, and stop for my
review. No code changes in this phase.

Phase 0 feels like a waste of a session. It isn't. When I ran it on my own rebuild, the audit discovered half the plan was already built in an earlier form, which changed the entire sequence. Ten minutes of reading saved weeks of duplicate work.

Step 4: Learn the checkpoint ritual

At the end of every phase the AI stops. Now you work. Read the checkpoint entry, click through whatever was built, then run the review prompt:

Phase [N] looks done. Before I approve:
1. List everything this phase changed.
2. List anything you left incomplete or worked around.
3. Run the app (or the build) and paste the result.
Then wait. If I say "approved", update PROGRESS.md, commit, and stop.
We start the next phase in a fresh session.

Push back freely at checkpoints. "The header is wrong, fix it before I approve" costs one message now. The same problem discovered three phases later costs a day.

Step 5: Start every session the same way

Every new session opens with the same two lines:

Read PROGRESS.md and MASTER_PLAN.md first. Then start the next phase.
One phase only, then stop for my review.

One phase per session keeps the AI's context clean, and the ritual makes progress boringly predictable: session, checkpoint, review, approve, repeat until the phase list is done.