The useful question is not whether MiroFish can generate text. It can. The useful question is how the system turns one input into a forecast you can actually audit.
Step 1: seed the scenario
Start with one report, policy note, launch brief, or narrative file. The source should contain actors, pressure, and a concrete question. Good inputs are narrow enough to model and rich enough to generate tension.
Step 2: extract entities and motives
MiroFish converts the source into structured nodes. Those nodes usually include institutions, individuals, audiences, media layers, and constraints. This stage matters because bad extraction creates false confidence later.
Step 3: build the graph
The graph is the operating surface. It makes the scenario visible instead of hiding everything inside one black-box response. You can inspect relationships, missing actors, and concentration of influence before running the simulation.
Step 4: run multi-round interaction
The simulation stage lets agent profiles react across Twitter- and Reddit-style surfaces. The point is not pixel-perfect realism. The point is to expose how a narrative might spread, fragment, or stabilize under pressure.
Step 5: read the forecast like an operator
The report should answer a bounded question:
- what changes first,
- who amplifies it,
- what risk becomes dominant,
- where confidence is weak.
Prompt template
Simulate how the uploaded brief evolves after the trigger event,
which actors gain narrative control, and what secondary effects appear
by the final round.
Related guides: What Is MiroFish? and LLM Social Simulation Explained.
Limits
If the source material is vague, the simulation will still look coherent. That is why scenario framing and review discipline matter more than raw token volume.