An Agent Does Work, Not Advice
Ask a chat how to make a landing page and you get a three-screen article. Give Claude Code the same sentence and minutes later a working page sits in your folder — it builds, checks its own result, and fixes it in a loop until the task is closed.
Why it works
A chat assistant hands you instructions and leaves the work to you; Claude Code does the work and hands you the result. Understanding this one difference changes how you phrase every request: you stop asking 'how do I…' and start saying 'build me…'. The payoff is measured in finished files, not saved articles.
How to do it
- Open the Claude desktop app and switch to the Code tab.
- Create an empty folder for the experiment and open it via Open Folder.
- Type a result, not a question: 'Build a one-page site in index.html with my name and three services.'
- Watch the working loop: it plans, writes the file, opens a preview, checks its own output.
- Review the diff it presents — green lines are additions — and click Accept.
- Open your folder in Finder or Explorer: the file is really there. That file is the difference.
Copy this
prompt
Build a one-page site in a single index.html: my name as the headline, three services, a contact line. Dark, restrained, looks good on a phone. Create the file and show me the result.
What to watch for
- If you phrase requests as questions ('how would I make a site?'), you get advice instead of files — the agent mirrors your framing.
- Don't dictate implementation steps; describe the finished thing and let it choose the how. You are the architect, not the typist.
Sources
Original: conon.aiDocumentation: code.claude.com