Five Ways to Pressure-Test an Idea with Synthetic Customers

MiroFish is an open-source multi-agent simulation project. It can generate many synthetic reactions to the same material. Those reactions are hypotheses, not real demand, survey data, or purchase predictions.

What this helps you do

Use one simulation to find unclear language, likely objections, conflicting interpretations, missing context, and questions worth taking to real customers.

Before you start

Use non-confidential material: a launch post, pricing explanation, ad draft, or product summary. Remove customer data and secrets. Review the repository, license, setup instructions, model costs, and data path before running it.

Terms to know

• Synthetic customer: an AI-generated persona, not a real respondent.
• Simulation: a model-generated scenario used to explore possibilities.
• Signal: a repeated pattern worth checking with real evidence.

Step-by-step

1. Open the official repository and read its current setup instructions.
2. Give the simulation one artifact and one narrow question.
3. Ask for distinct reactions, objections, and disagreements rather than one average score.
4. Group repeated concerns, then trace each concern back to the text that prompted it.
5. Turn the strongest patterns into interview questions or a small real-world test.

Copy this

Material to review: [PASTE OR DESCRIBE THE LAUNCH POST, PRICE, AD, OR PRODUCT]

Question: [ONE THING YOU NEED TO LEARN]

Generate varied synthetic reactions. Separate:
1. what was clear,
2. what was misunderstood,
3. likely objections,
4. missing information,
5. disagreements between personas.

Do not call this customer research or predict purchases. End with five questions I should ask real customers.

Things to know

More personas do not make the result statistically representative. The output inherits the model's assumptions and the framing in your prompt. Use it before real research, not instead of real research.

If something goes wrong

• If every persona agrees, ask for conflicting interpretations and edge cases.
• If the feedback is generic, narrow the audience and include the exact artifact.
• If setup requests credentials you do not understand, stop and review the repository before continuing.

Final check

Pass only when you have five testable questions for real people and none of the synthetic output is presented as validated demand.