Treat it like a junior, not a tool

Claude Code got more useful when I stopped treating it like autocomplete and started giving it design feedback.

I got mediocre results when I treated AI like autocomplete: prompt, output, cleanup, repeat. The work improved when I started treating the agent like a junior designer. I gave it specific feedback, pointed to examples, and made it try again.

Concrete observation beats abstract direction

A junior designer told to make it more elegant produces nothing useful. Same is true of the AI. Tell it the spacing under the button is twelve pixels and you want eight, you get eight. Tell it the gallery is feeling cropped and you want the natural image size preserved, you get the right CSS. Be specific about the bug, not the vibe. The model converges fast on a named problem. It flounders on a named feeling.

Use git as the undo button

Working with a junior, you accept that some directions will be wrong, and you have a lightweight way to roll them back. Same here. Commit before any structural change. Push to a branch before any speculative redesign. The agent is more useful when you are willing to throw away its work than when you are protecting it. This looks worse than before, put it back is a complete instruction in a working session, not a failure mode.

Critique by example, not by principle

Match the density of this list. Match the type scale on that page. Use the chrome from this component. The model is excellent at copying a real artifact in the codebase and bad at internalizing a stated rule. Show it where the answer already lives. The same trick I use with junior designers turns out to work on the agent.

Tell the agent where the answer lives

The agent has a particular way of failing. It handles a named problem better than a vague direction, and a working example better than a principle. I get better work when I make changes reversible and point to the part of the codebase that already has the answer. The model did not change. My feedback did.