From an idea to a build you understand.

Start with the job, not the screen.

Write down who the product is for, what they need to do, and what a successful result looks like. A clear task gives you something to judge the first output against.

For a background tool, success might mean: choose a format, change the surface, add a title, and download the image. The visual style should support that sequence.

Give the AI the right context.

Include the relevant project files, existing conventions, examples and constraints. Ask it to inspect the current implementation before changing it. Explain what should stay intact, especially work already published or used elsewhere.

Build one complete interaction.

Ask for a narrow change that works end to end. “Make the export button produce a PNG in the selected size” is easier to verify than “finish the whole platform.” A small working slice reveals gaps early.

Inspect what actually changed.

Read the explanation and review the changed files. Run the main interaction yourself. Test with keyboard and touch where relevant, and check an error or empty state. A confident response is not a substitute for a working result.

Describe the mismatch precisely.

Explain the action, expected result and observed result. Include the relevant error or screenshot. Ask for the cause and a focused repair before requesting another broad redesign.

Check the published version.

A saved file, a Git commit and a successful deployment are different checkpoints. Open the published version and verify the same interaction you tested locally. Record the version used by any tutorial or demonstration.

Try it

Open Phase Memory, choose one small improvement you would make, and write a brief for it. Include three checks that would prove your change works. Use the planning prompt in the library to turn that brief into a bounded implementation task.