Build it step by step

01What this module is about

You have a project set up, a clear way of asking, and a plan. Now you build.

The biggest mistake at this stage is asking for everything at once. "Write the whole newsletter, the poster and the email to members" gets you three average results. And when something is wrong, you can't tell where it went wrong.

This module gives you a simple loop: build one piece, check it, save it, then move on. It feels slower at first. It saves time later, because mistakes get caught early.

By the end you'll have built the first real piece of your project, checked it with fresh eyes, and saved a good copy.

02Why small pieces work

• You can check each one. A short piece takes minutes to check. A huge one never really gets checked.
• A mistake costs one step, not everything. If a piece goes wrong, you redo that piece. The rest is safe.
• The AI does better with one clear job. One request, one result. OpenAI's guide for everyday work gives the same advice.

03The loop

1. Pick one step from your plan.
2. Build it in a fresh chat inside your project.
3. Check it in a different fresh chat.
4. Save the good version, then pick the next step.

A fresh chat for each step keeps the AI focused on that step, as you saw in Module 4. Anthropic recommends the same for Claude Code. Write the plan first, then start a fresh session to build from it.

04Text to copy

To build one step, start a fresh chat inside your project and paste this:

Read PRD.md and this step from my plan:
[paste the step]

Build only this step. Stop when it's done.
When you finish, tell me how you checked it.

Your starter lines from Module 1 already tell the AI to ask if anything is unclear, and to stop if it gets stuck. So you don't need to say it again here.

To check it, start another fresh chat and paste this:

Here is a piece of my project, and the step from my plan it's meant to do.
Check it against PRD.md and that step.
List only problems that affect whether it's right or does the job, most serious first.
Skip style and wording preferences.

Step: [paste the step]
The piece: [paste it]

The last line matters. Anthropic warns that a reviewer asked to find problems will always find something. If you fix every small comment, you end up making things worse. Fix what affects whether it's right.

06When it goes wrong

• It's going round in circles. After two tries that don't fix it, stop. Start a fresh chat with your handover note and a clearer request. More tries in the same chat rarely help.
• It changed things you didn't ask about. Say exactly what to change: "Only change the meeting time. Leave everything else as it is."
• The piece is too big to check. Split the step in two, and build each half separately.
• It says it's done, but it isn't. Give it a way to check, and ask it to show the check. For example: "Check every date against the Reference file and list them."

07Things that don't help

• Telling it to keep going or not be lazy. Anthropic says lines like that are more likely to cause over-work than help. If it stops short, just say "continue".
• Fixing every review comment. Some comments are only taste. Fix what affects whether it's right.
• Keeping one long chat for the whole project. Use one fresh chat per step, with your project files and handover note to carry what matters.

08Keeping it current

• Tick off each step on your plan as it's done.
• If the plan changes, update PRD.md first, then carry on.
• Keep your version files. If a later change goes wrong, you can always go back to the last good one.

09Keep building

You don't need to finish the whole project before you move on. Carry on to Module 7 with your first piece.

Keep using this loop for the rest of your plan as you go through the next modules. Plan, build, check, save. Each module from here adds a habit you can use on every piece.

When you reach Module 10, it's fine if only part of the project is done. You can run Module 10 on what you've finished so far.