Claude Code for Non-Developers: What to Set Up First

You followed a beginner guide and it worked. Claude Code is open in the right folder, there is a CLAUDE.md at the top of the project, and the first afternoon produced a sign-up page that actually signs people up.

Three weeks later you open a new session and ask for a pricing page. Claude Code makes a confident plan and starts rebuilding the account screen you settled in week one, because nothing it can see says you settled it.

The setup was fine. It covered the first afternoon, and you are somewhere else now.

What the beginner guides get right

Most guides written for non-developers agree on the basics, and they are right. Open Claude Code from the project folder. Write a CLAUDE.md that says what the project is and how you like to work. Commit to git before each task, so a bad session is one undo away. Use plan mode, so Claude Code shows you what it intends before it touches a file.

Do all of that. It is the difference between a frightening first week and a fun one.

What they leave out is the project that lasts. A real project is the one you are still working on in week four, after you have changed your mind twice. That is where non-developers get stuck. Not on syntax. On keeping the thing pointed where they meant it to go.

Why week four feels different from day one

Claude Code is very good at the next change. It does not plan the project, because the project lives in your head and in old conversations, and it reads neither.

Conversations also end. You start a fresh one when the current one gets long or strange, which the guides rightly recommend, and whatever you decided in the old one stays there.

The CLAUDE.md helps, up to a point. Files describe. They don't decide. It says how to build, but nothing checks finished work against it, and it only changes when you remember to change it.

None of this means you are doing it wrong. Your idea grew. Ideas do that once they meet real users, and changing your mind is the project growing in scope, not a lack of discipline.

Write the project down before you build it

Before the first real task, write a few plain pages. None of them need code.

A one-page brief. What the product is, who it is for, and what is deliberately not in the first version. The last part matters most, because Claude Code will happily build anything you have not ruled out.

A short ordered list. The next 3 to 8 tasks, in the order you want them. Just the next stretch of work, small enough to finish and look at.

A decisions page. Each time you settle something that matters, write one line on what you chose, why, and how sure you were. When you change your mind, add a new entry and keep the old one. We wrote about why that history matters in Architecture Decision Records Do Not Survive AI Speed.

Then ask Claude Code to read those pages at the start of every session, and to name the task it is on before it begins.

Define done in words you can check

This step matters most if you cannot read code. For each task, write down how you will know it is finished, in terms you can test by clicking around.

Not “add accounts.” Instead: I can sign up with an email, log out, log back in, and my saved project is still there. A wrong password shows a message, not a blank screen.

You do not have to read the code to judge the work. You have to decide what done means before the work starts.

After each task, check it against that list yourself. Then write two lines: what surprised you, and anything broken you noticed on the way. Read them before you plan the next task.

Running it without a second job

Kept by hand, those pages drift, because updating them competes with building. That gap is what PAPI is for.

PAPI is a project manager that works inside your AI chat. It holds your vision, your decisions and your progress as your idea grows in scope, so you stay in control while you build. In Claude Code you ask it to plan, and it turns the next stretch of work into a short cycle where each task gets a build handoff: the goal, what is out of scope, and how to check it is done. Every build files a report. You give each task a verdict: accept, request changes, or reject. Decisions keep their reasoning and are superseded, never overwritten.

How it connects is on the Claude Code page.

The honest version

PAPI does not write code, and it does not make Claude Code any smarter. Your assistant still does the building, on the subscription you already pay for.

If you are still exploring, still throwing away most of what you build, skip all of this, PAPI included. A notes file is plenty. The setup above earns its keep the first time you catch yourself explaining the same project to the same assistant again.

To see a planned project before you start your own, walk through the live demo, no signup needed.