Scenario guide

Pastura runs multiple AI agents on your device. Scenarios are written in YAML, and a local LLM plays each agent in line with the scenario’s personas. This page walks through what a scenario looks like and how to write one.

How a scenario works

A scenario is a single YAML file. It declares how many agents take part, who they are, and what they do each round. Three pieces carry most of the meaning:

agents and rounds set the size of the scenario.
personas give each agent a name and a short character brief.
phases describe what happens in order, such as everyone speaking, making a choice, voting, or tallying a score.

You can write that file by hand, or build it in the app’s visual editor instead. Either way, the app ships with several presets, so you can watch a run before writing one yourself.

A worked example

The Prisoner’s Dilemma preset puts five contestants through a classic strategy game. Each round they address the group, then choose to cooperate or betray each opponent in turn. A scoring phase converts those choices into points, and bluffing during the spoken phase is fair play. Here is the shape of it in YAML, with the prompt text left out:

name: Prisoner's Dilemma
agents: 5
rounds: 3
context: >
  You are a contestant on the game show "Prisoner's Dilemma".
  Against each opponent you choose "cooperate" or "betray".
  Both cooperate = 3 points each. You alone betray = 5 points.

personas:
  - name: Alex
    description: >
      [Role] Calm strategist.
      [Goal] Compute the optimal move from data and past behaviour.
  - name: Mia
    description: >
      [Role] Optimistic people-pleaser.
      [Goal] Trust people by default, even after being burned.

phases:
  - type: speak_all      # everyone addresses the group
    output:
      statement: string
      inner_thought: string
  - type: choose         # pick cooperate or betray, per opponent
    output:
      action: string
      inner_thought: string
    options:
      - cooperate
      - betray
    pairing: round_robin
  - type: score_calc     # built-in prisoners_dilemma scoring
    logic: prisoners_dilemma
  - type: summarize      # recap the round's scoreboard
    template: >
      Round {current_round} result: {scoreboard}

Five personas, three rounds, four phases. The persona briefs are where the run gets its character. A calm strategist and an optimistic people-pleaser will say very different things on the same model.

More presets

The other presets each highlight a different mechanic.

Scenario	Agents	Phases	What you can watch for
Word Wolf	5	9	Five players get a topic word, but one of them gets a different word. Watch how the odd one out hides, and whether the group exposes them in the vote.
Bokete	5	5	Five comedians caption a photo prompt, then vote on the funniest bit. Watch how each persona’s comedic style holds up under a crowd vote.
Score Race	3	4	A vote-driven points race to a target score. The round-end summary changes once someone reaches the target, so you can see how a phase branches on game state.

Shared Scenarios

The presets are just the doorway. The app’s Browse tab holds a larger collection called Shared Scenarios, ranging from cognitive-bias experiments to word games and offbeat role-play settings, each with a different angle from the presets.

One example is The Pasture Council of Sheep, where five sheep on a shepherd-less farm debate whether to step beyond the fence. It grew out of a finding that a strong persona brief lets even a small local model keep each character distinct. You can look inside one from its share page.

Writing your own

Every preset is a starting point. The editor has a visual mode, where names, personas, and phases are form fields, and a YAML mode for direct control over the file shown above. Open a scenario in either mode, change a persona or a phase, and run it again to see how the outcome shifts. A few things are worth doing as you experiment:

Run the same scenario several times. Because the AI is non-deterministic, its output shifts every run, so a single run won’t tell you much about the scenario. Look for patterns across runs.
Vary a persona’s description. Adding a trait like “confident” or “eager to please” can change how the run plays out.
Reveal the inner voice. It shows what was behind an agent’s words.

For the full YAML format, every phase type and output field, the scoring logics, and the common pitfalls, see thescenario format reference. A raw Markdown version lives at format.md, handy for handing to an LLM.