What is counterfactual reasoning?

It is asking what would have happened if one condition were different, holding everything else constant. Instead of predicting the future, it re-runs a known scenario with one altered assumption to isolate that factor's true effect.

Why change only one variable?

Changing several variables at once makes it impossible to attribute the new outcome to any single cause. The prompt enforces a single change so the resulting causal chain stays interpretable and the analysis is honest.

What does reasoning depth control?

It sets how many causal hops the model traces from the changed variable. Shallow depth gives an immediate effect, while deeper settings follow second- and third-order consequences and any feedback loops.

When is this kind of prompt useful?

Post-mortems ("if we had shipped a week earlier"), strategy ("if the price were halved"), debugging ("if that config were off"), and history or fiction. Anywhere isolating one factor's effect adds clarity.

Will the model just guess?

The generated prompt asks the model to state its assumptions, reason step by step through the causal chain, and flag low-confidence links. That structure surfaces guesses as explicit, labeled uncertainty rather than hiding them.

Counterfactual Reasoning Prompt Builder

Counterfactual reasoning prompt builder

Counterfactual reasoning is the disciplined version of “what if” — you take a known situation, change exactly one assumption, and trace what would follow. It is how analysts isolate cause from coincidence and how good post-mortems avoid hindsight bias. This builder produces a prompt that forces an LLM to do this properly: hold everything constant, flip one variable, and follow the causal chain step by step rather than leaping to a tidy conclusion.

How it works

You provide three things: the baseline situation, the single variable to change, and how deep to reason. The tool wraps these in a prompt that instructs the model to first restate the baseline and its key assumptions, then apply the one change, then trace consequences hop by hop — each step naming what it causes and why. Crucially the prompt tells the model to keep all other conditions fixed and to label any link it is unsure about, so speculation is visible instead of smuggled in as fact. Deeper settings ask for second- and third-order effects and feedback loops; shallower settings stop at the immediate consequence.

Tips and example

Change one thing. If you want to compare several alternatives, generate one prompt per alternative rather than asking the model to juggle them — comparisons stay cleaner and the causal chains stay attributable. A good example: baseline “we launched the feature with no onboarding flow and saw 20% activation,” variable to change “we added a three-step onboarding flow,” depth “trace second- order effects.” Watch for the model’s flagged uncertainties — those are exactly the assumptions worth validating with real data before you act on the analysis.