Plan first, then act

So far Claude has only written text back to you. Give it tools (a web search, your files, a terminal) and a prompt stops being a question. It becomes a job Claude does: it searches, reads, edits and runs things, one step after another.

That raises the stakes on one small habit. Before a job with several steps, ask for the plan, and stop there. Read it, fix it, then say go.

The reason is simple: a wrong plan costs more once it's running. If Claude misreads the job in step one, every search, edit and command after it builds on the mistake. A plan is cheap to read and cheap to change.

The second half of the habit is to say what you want done, in plain verbs. Anthropic's guide says its latest models are trained to follow instructions precisely. Ask "can you suggest some changes?" and Claude will often give you suggestions instead of making them. Write "change this function to…" and it makes the change.

The left reply is vague because the prompt was. The right one came from a prompt that said exactly what to produce, how long, and when to stop.

A good plan-first prompt has four parts:

1. The job, with the detail that matters ("keep every old URL working").

2. Plan only: a numbered list of steps, with a limit, say six.

3. A stop: "don't start yet" or "wait for my go-ahead".

4. The question back: ask Claude to end by checking the plan with you.

Coding agents build this in. Claude Code has a plan mode, and the Work with coding agents course uses it. But the habit works with any tool, and with no tools at all, which is how you'll practise it here.

Your turn. The job is the blog move from the example. Write a prompt that asks for a numbered plan of at most six steps, says don't start yet, and asks Claude to finish by checking the plan with you.