AI Guides › Claude Mastery

The One Instruction That Fixes Most Bad Outputs

By Nigel Guy · 2 min read

When an answer comes back wrong, most people's first move is to correct the output — rewrite this paragraph, change that number, try a different angle. That treats the symptom. It's slower than it looks, because you're debugging the finished answer instead of the thinking that produced it, and the same misunderstanding tends to resurface two edits later wearing a different sentence.

The mechanism

1. Add one line to your request: "Before you answer, tell me your plan and what you're assuming." This single instruction is doing almost all of the work in this guide — everything else is detail on using it well.
2. Read the plan for the assumption, not the polish. You're not checking whether the plan sounds competent — you're checking whether it's solving the problem you actually meant. A confidently-written plan built on the wrong assumption is still wrong.
3. Correct at the plan stage, in one sentence. "That's the right structure, but assume a technical audience, not a general one" costs you a sentence. The same correction after a full draft costs you a rewrite.
4. Only then ask for the full answer. The plan step isn't extra work bolted onto the task — it replaces a round of "no, not like that" after the fact with a much cheaper round before it.
5. Drop the step once a pattern is established. For a task you've run the same way several times with the same person or Project context, the plan step has already done its job — you don't need to re-run it on every single instance of a task you've already calibrated.

What to skip

Skip asking for the reasoning on something genuinely simple — a one-line factual question doesn't need a plan, and asking for one is its own kind of theatre. And skip treating the plan as a formality to approve without reading; if you're rubber-stamping the plan step the same way you'd have rubber-stamped the finished answer, you've kept the extra step without keeping the benefit.

Guardrails

• This fixes misunderstandings about what you actually want, which is the majority of "bad outputs" in ordinary use. It doesn't fix a genuinely wrong fact buried inside an otherwise well-aimed plan — that still needs your own verification, the same as any finished answer would.
• A stated plan is not a guarantee the final answer will follow it exactly, particularly on longer or more complex tasks. Treat it as a strong steer, not a contract.
• This works better the more specific your correction at the plan stage is. "Looks good" as a response to a plan gives you the same odds as skipping the step entirely — the value is in actually naming what's wrong, not in the extra exchange itself.

All 751 AI guides · JulieMango plans from £17/mo