The Show Your Work Trick That Catches AI Errors Before You Do

The most expensive AI errors are the ones the model delivers confidently and you accept without checking. The fix is not to spend more time fact checking after the fact. The fix is to force the model to show its work before it commits to an answer, so the weak reasoning surfaces before you can be fooled by the confident delivery.

The trick

Add this to any prompt where the answer matters:

Before giving the final answer, walk through your reasoning step by step. State every assumption you are making explicitly. Then give the answer.

That is the entire intervention. Forces the model to expose its thinking. The errors and weak assumptions become visible where you can catch them, instead of hiding inside a confident answer.

Why it works

The default AI response delivers the conclusion first because that is what most users want. The problem with conclusion first is that the reasoning is hidden behind it, and confident phrasing covers for weak logic. The show your work instruction inverts the order. The reasoning comes first, the conclusion last. You read the reasoning before you read the conclusion, and any weak step becomes visible.

When to use it

Use it on any output that involves calculation, multi step logic, comparison, decision support, or anything where being wrong has a cost. Skip it on creative drafting or exploratory work where the model's intuitive jump is the value.

Three sharper variations

For calculations: "Show every calculation step. State the inputs you are using and where you got them. Flag any number that is an estimate."

For analysis: "State three assumptions you are making before you analyse. Then do the analysis. Then state which of the three assumptions, if wrong, would most change the conclusion."

For research: "List the sources or knowledge you are drawing on for this answer. Rate your confidence in each. Then give the answer, with the confidence rating attached."

Before and after

Without the trick, asked to assess whether a pricing model is sound: "Your proposed pricing of $99 per month for the mid tier is competitive and well positioned given the market context."

With the trick: "Reasoning: I am assuming (1) your competitors are in the $79 to $149 range, (2) your customers expect tiering with mid tier as the anchor, (3) your COGS at this price leaves 70 percent margin. If assumption 1 is wrong (e.g. competitors are actually $49 to $79), the pricing is too high. If assumption 3 is wrong (e.g. margin is closer to 40 percent), the pricing is too low for unit economics. Conclusion: pricing looks sound if the assumptions hold. The weakest assumption is the competitor range, which I would verify before committing."

The second output is what you actually needed. The first one would have walked you into a wrong decision while you nodded along.

The deeper version

The full system for AI that catches its own errors before you do is the CRITIC Framework: fifty prompts across six pillars of adversarial thinking. The show your work trick is the entry point. CRITIC is the toolkit. Pair it with the AI Workflow Audit to identify where in your week the trick will save you the most.

The reframe

The model is not lying to you. It is committing to answers based on weak reasoning that you cannot see because the reasoning is hidden behind the confident delivery. One instruction makes the reasoning visible. That visibility is the entire difference between AI that augments your judgement and AI that occasionally walks you off a cliff while you nod.

PromptLeadz publishes battle tested AI prompt packs for operators across all functions. All prompts are LLM agnostic. Pricing is in USD.