Give It Instructions, Not Code

The single thing that decides whether Claude Code nails a task is how you ask. Good news: the fix has nothing to do with code. It's just clearer English.

Here's a vague instruction:

clean up my photos

And here's the same job, briefed well:

In the Photos folder, rename each file to the date it was taken plus a short description, like 2026-03-birthday-cake.jpg. Don't touch anything in the Edited subfolder, and ask me before renaming if you're unsure about a date.

The second one works far better, and not because it's fancier. It works because it says three things the first one didn't: what "clean up" actually means, what to leave alone, and what to do when unsure.

That's the whole skill. What you're writing is called a prompt, which is just your instruction to the assistant in plain words. A good one usually has:

• One job, not five. Ask it to rename files, then ask about deleting duplicates after. One task at a time keeps it focused and keeps you in control.
• A clear "don't touch this." Naming what's off-limits prevents the mistake you'd most regret.
• Room to ask. Telling it to check with you when unsure beats it guessing wrong.

If a task is big, you don't have to plan every step yourself. Add think through this step by step before you start and Claude will work out the order first. See thinking step by stepsoon for why that helps.

One more thing: Claude only knows what's in the current conversation and the files it can see. It doesn't remember last week. If a detail matters, put it in the instruction. That working memory has a size limit called the context windowsoon.

Try this now: take a task you'd normally rush, and write it out with all three parts above before you paste it. Compare the result to your usual one-liner.

Where to go next: Permission Modes and How to Undo a Mistake covers what happens when an instruction doesn't go the way you meant. For more ways non-engineers phrase their requests, see MindStudio's Claude Code patterns for non-engineers.