How to Write a Prediction Prompt: 7 Rules That Transform Your Reports

A prediction is only as good as the scenario you feed it. MiroFish enriches every prompt automatically, but the difference between a generic report and one that changes your decision usually comes down to seven things you control.

Rule 1: Name a specific actor

"A company launches a product" simulates a stereotype. "A 15-person bootstrapped SaaS with 2,000 paying customers" simulates something real.

• Weak: "What happens if a brand raises prices?"
• Strong: "What happens if Basecamp-style project tool with 8,000 SMB customers raises prices 20%?"

Rule 2: Describe one concrete event

Simulations need a stone hitting the water. Multiple simultaneous events blur the ripples.

• Weak: "We're rebranding, raising prices, and launching a new tier."
• Strong: "We're raising the Pro tier from $29 to $39 on March 1." (Run the rebrand as a second prediction.)

Rule 3: Say who you care about

The report will prioritize the stakeholders you name. If churn among agencies is what keeps you up at night, say so.

Rule 4: Set a time horizon

"First 30 days" produces different dynamics than "over two years." Short horizons surface reaction; long horizons surface adaptation.

Rule 5: Include the constraint that makes it hard

Every interesting decision has a tension. Give the simulation the real trade-off:

"We need the revenue from the price increase to fund support hiring — but our biggest accounts are the most price-sensitive."

Agents will fight about exactly that tension, which is the fight you need to watch.

Rule 6: Attach the document, not the summary

If a launch brief, positioning doc, or policy text exists, attach it. Ten pages of specifics beat your three-sentence memory of them. (MiroFish accepts .txt and .md uploads.)

Rule 7: Ask for the decision, not just the weather

End the prompt with the choice you actually face:

• Weather: "Predict the reaction to our launch."
• Decision: "Predict the reaction to our launch — and whether we should lead with the free tier or enterprise pilot program."

A full example, assembled

"Lumen, a 12-person indie email client with 40k free users and 3k paying ($8/mo), will introduce an AI inbox-triage feature as a $4/mo add-on on May 1. Core users are privacy-conscious developers and writers; a loud minority is hostile to AI features. Predict reaction over the first 60 days across our subreddit, Hacker News, and tech press — and whether bundling the feature into the existing paid tier would produce a better outcome than the add-on."

Actor, event, stakeholders, horizon, tension, decision — six rules in one paragraph (the seventh is attaching the positioning doc).

Structured prompting like this is the same discipline good forecasting research demands — Philip Tetlock's superforecasters outperform precisely because they decompose vague questions into specific, scoreable ones.

The three inputs that reliably waste a run

Vague prompts do not fail loudly — they produce a confident, generic report, which is worse. The patterns that cost people a credit:

No incumbent. A concept described without what it replaces gets an unrealistically warm reception, because the simulated evaluators have nothing to compare it against. Name the thing people use today, even if that thing is a spreadsheet or doing nothing.

No named segments. "Our users" is not a segmentable input. Naming even two or three real groups changes the output more than any other single edit, because the engine reports where their reactions diverge rather than averaging them into mush.

No stated tension. If you do not say what makes the decision hard, the report will not find the hard part for you. The fear you would rather not type is usually the most valuable sentence in the prompt.

There is a fourth, subtler one: asking for a number the engine cannot know. Derived figures are the weakest part of any report — our own published example computes a $6,000 loss where the arithmetic gives roughly $2,000, and we left it visible for exactly this reason. Ask for reactions and mechanisms; do your own arithmetic.

Browse ten more worked examples, grab a starting point from the prompt library, or test your prompt live.