My Claude Code workflow, step by step

People assume building with AI means typing “make me an app” and praying. It doesn’t. It’s a discipline, and the discipline is most of the value. Here’s the loop I actually run.

1. Plan before a single file changes

I never let the AI touch code on the first message. I make it read the project, propose an approach, and list the risks first. If the plan is wrong, the code will be wrong — and code is far more expensive to throw away than a paragraph.

Cheap to change a plan. Expensive to change a codebase. Spend your arguing-time at the stage where it’s cheap.

2. Work in small, verifiable steps

The single biggest upgrade to my output was forcing one change at a time. After each step, I stop and check it in the browser before moving on.

# After every meaningful change, the same three commands.
npm run dev      # see it working locally
npm run build    # catch anything that breaks the real build
git add -p       # review each change before it's mine

If I can’t verify a step, the step was too big. I make the AI split it.

3. Read the diff like it’s a contract

This is the part non-coders skip, and it’s the part that matters most. I don’t need to be able to write the code. I need to be able to read it well enough to catch when it’s doing something I didn’t ask for.

A metrics dashboard used as an example of verifying output against expectations — I check the output against what I expected — not against what the AI claims it did.

The traps that cost me afternoons

Letting it “fix” five things at once, then not knowing which fix broke things.
Accepting a confident explanation instead of checking the actual result.
Skipping the build step and discovering the breakage after I’d moved on.

4. Keep a record of what worked

When a prompt or approach works well, I save it. The workflow compounds — month three is dramatically faster than month one, purely because I stopped re-solving the same problems.

That’s the whole playbook. Not magic. A loop, run with discipline.