Skip to main content

Superpower techniques

The prompting formula shapes a single request. These are the habits that stop Claude solving the wrong problem. Use them before anything that matters.

Start with the problem, not the solution​

Don't open with the deliverable. Open with the problem.

The problem I have is X. If I were to get you all the data you need, can you help me think through the problem and plan, without executing? Only once we understand the problem deeply should you go into the solution.

This is the same technique a consultant uses. A good consultant working for you would understand all the context around the company, and start with the problem, before jumping to any solution. It's exactly the same with Claude.

Give it the context. The files in your folder are that context, once the folder is selected. See Context is king.

Plan before you execute​

When the problem is clear, get a plan. Don't let it edit a file, send a message, or build anything until you've read the plan and corrected it.

Show me what you'll do. Don't change anything yet.

Read it. Correct it. Then tell it to go.

Ask me any clarifying questions​

Tell it to ask. That cuts down assumptions, so Claude (or ChatGPT) makes less up, because it has the information it needs.

Ask me any clarifying questions before you start.
Don't fill a gap with a guess.

On Claude it can ask one at a time. Answer them. A short prompt plus a few good answers beats a long brief you typed in a hurry.

Plan anything complex​

Decks, financial models, and of course skills and scheduled tasks. Plan them before anyone builds them.

Plan on Opus, or on Fable when the job is rare and being wrong is expensive. If you're watching your usage, do the execution on a lower model once the plan is agreed. See Models, effort and limits.

Plan this in text first. Headings, the point of each part, and what you still need from me.
Don't build the file yet.
Ask me any clarifying questions first.

When the plan is right: "Build this exact plan."