How to write a scenario question to get a useful prediction

The biggest factor in how useful a prediction is isn't the model — it's the question. A scenario described in one fuzzy sentence forces the AI to invent your entire situation, and the prediction is then mostly a prediction about the AI's guesses. A scenario described well gives the model real material to reason from, and the prediction sharpens accordingly. This is the highest-leverage skill on the whole site, and it takes about two minutes to learn.

Here's how to write a scenario question that earns a sharp prediction, with concrete before-and-after examples.

Rule 1: Name the outcome you care about

Vague questions hide the outcome. "Will my business do well?" — do well by what measure? Revenue? Survival? Your sanity? The AI has to pick one, and it might pick the wrong one. Sharp questions name the outcome explicitly.

• Before: "Will raising prices be good for us?"
• After: "Predict the churn-adjusted net revenue impact over 12 months if we raise prices 15%."

The "after" version tells the prediction exactly what to optimize and report. (You can see this play out fully in predicting customer reaction to a price change.)

Rule 2: Quantify whatever you can

Numbers are the difference between a prediction grounded in your reality and one built on the model's defaults. Every number you supply is one fewer thing the AI has to guess.

• Before: "We have some customers and we're thinking of raising prices."
• After: "We have 800 mid-market customers at $8k annual ARPU and are considering a 15% increase."

You don't need precision — rough, honest numbers beat none. "About 800 customers, give or take" is enormously more useful than no figure at all.

Rule 3: Include the awkward context

The details people leave out because they're embarrassing or complicated are often the most decision-relevant. A prediction can't account for the risk you hid from it.

• Before: "Predict the outcome of hiring this senior engineer."
• After: "Predict the outcome of hiring this senior engineer — they're brilliant but a known culture risk, and they'd join a flat team with no dedicated manager."

The awkward clause is the whole prediction. Without it, you get a generic answer; with it, the prediction can find the variable that actually decides the outcome (see predicting the outcome of a hiring decision).

Rule 4: State the rival, not just the plan

The most useful predictions compare specific alternatives. A scenario with no alternative forces the AI to invent the counterfactual.

• Before: "Predict what happens if I invest my savings."
• After: "Predict my net worth in 10 years if I max retirement contributions vs. paying down my 6% mortgage early."

Two named options give a clean comparison; one vague intention gives a vague forecast. (Predicting your financial future leans entirely on this.)

Rule 5: Set the horizon and let it ask

Tell the prediction how far out you care about — 90 days, a year, five years — because reliability and the right reasoning both depend on the horizon. And if you've left a gap, let the AI ask before it predicts. A good prediction tool will request the missing context rather than fabricate it; answering one or two follow-ups is faster than re-running a prediction built on bad guesses.

Putting it together

A strong scenario question names the outcome, quantifies what it can, includes the awkward context, states the rival option, and sets a horizon. Compare the two ends of the spectrum:

• Weak: "Should I change jobs?"
• Strong: "Predict the outcome over two years if I leave my $120k senior role for a smaller company at a 20% pay cut with meaningful equity, where the scope is described loosely — given that my current role is stable but stalled."

The strong version gets a prediction you can act on. The weak version gets a prediction about a job change the AI mostly imagined.

This is the skill that compounds across everything else — it's why some predictions are more reliable than others and the input side of how AI scenario prediction actually works. Spend the extra two minutes on the question. It's the cheapest way to make every prediction better.