How do I stop AI charging off in the wrong direction?
It's much cheaper to fix a plan than a finished job.
You ask Claude Code to "tidy up the monthly report script".
AI gets to work. It renames things, reorganises the folders and rewrites half the logic. Forty minutes later it's done.
You only wanted the column names fixed.
Made-up example.
A vague request has lots of reasonable readings. It picks one and goes. It doesn't stop to check which one you meant unless you ask it to.
And once it's started, every step it takes is on the desk, nudging the next step in the same direction (see "Why AI gets stuck on its first idea").
So the time to steer is before the first step.
Before you change anything, tell me your plan: what you'll change, in what order, and anything you're unsure about. Wait for my OK.
Now you're reviewing three lines of plan instead of forty minutes of work.
Most people don't do this. In "AI can't read your mind" we saw it in only about 1 in 10 conversations.
Some tools build this in.
Plan mode: a setting in Claude Code where it can read and think, but isn't allowed to change anything. It writes up a plan and waits for you to approve it before it starts.
It's like asking a builder for a quote and a drawing before they pick up a hammer.
When you're not sure what to ask for, flip it round:
I want to write a proposal for a four-day week trial. Ask me questions, one at a time, until you have what you need. Then write it.
AI will ask the questions a good colleague would: who's it for, what's the budget, what are people worried about.
You'll often find the questions are more useful than the first draft would have been.
Made-up example.
Not everything needs a plan. Quick questions, a rewrite of one paragraph, "what's the formula for this": just ask.
It's worth it when:
For anything big, ask for the plan first and wait to approve it.
In Claude Code, use plan mode for anything that touches more than one file.
When you don't know what you want, ask it to interview you.