| name | simulation-design |
| description | Use this skill in the DRF-2 pipeline to design the multi-agent OASIS simulation from the actor dossier and knowledge graph — dossier-grounded persona design (identity/views/incentives/relations/red-lines), the shared world brief, seed posts from researched stances, activity profiles and round mapping, scheduled events from the dated timeline, and the initial follow graph with valence-aware clustering. The output is the simulation_config.json contract consumed by the simulation engine's sim_start tool. |
| allowed-tools | ["tool_search","read_file","write_file","ls","grep","glob","kg_search","kg_get_entities","kg_n_hop_subgraph","kg_centrality_priors","sim_start","sim_status","sim_results","sim_interview_agents"] |
Simulation Design Skill
1. Mission & prime directive
Turn the researched cast into a grounded multi-agent social simulation — not a generic role-play. Every design choice below exists to defeat the "hollow sim" failure mode, where agents with thin personas free-associate instead of behaving like the real actors.
Prime directive: everything is dossier-grounded. Personas, stances, seed posts, follows, and events are distillations of researched evidence (the actor dossier, actors.json, the knowledge graph) — never inventions. Cognitive diversity comes from the real differences between real actors, not from artificially assigned analytical schools. If the dossier does not support a stance, the persona does not have that stance.
Inputs: actors.json (cast + relationships + situation brief + key events + forecast inputs + hot topics), the KG (kg_* tools for ego networks and centrality), and the run's round budget. Output: the fields of simulation_config.json (§7).
2. Persona design (per selected agent)
Each cast-backed agent gets a structured persona_design object with eight fields, each distilled from the dossier's evidence for that specific actor:
| Field | Content | Grounding source |
|---|
identity | Who it is; its role position in the system | actor role + worldview.identity |
views_beliefs | Its worldview and reading of the current situation — what it believes is true | stance / beliefs / framing |
incentives | Gains/losses under each plausible outcome; the objective function it optimizes | incentives (gains-if / loses-if) |
objectives | Goals and its strategic path to them | goals + strategy |
relations | Allies / opponents / dependencies and WHY — named counterparties | relational roster + relationships[] (valence preserved) |
constraints_red_lines | Hard constraints and lines it will not cross | constraints + vulnerabilities |
decision_style | How it weighs trade-offs; its action tempo | risk tolerance / track record |
rhetoric | Its typical public voice and argumentative style | stated positions, quoted language |
Rules:
- No invented stances or facts. Every field must trace to dossier evidence; the stated-vs-revealed gap, when the dossier surfaces one, belongs in
views_beliefs/decision_style.
- The persona prose (the agent's system-prompt biography) must embody the persona_design, and both must use the actor's canonical name.
- Enrich
relations with the actor's KG ego network: retrieve 1–2 hop neighbors filtered to relationship edges (ALLY_OF, OPPOSES, COMPETES_WITH, REGULATES, DEPENDS_ON, PARTNERS_WITH, INFLUENCES) so the persona knows who to @, ally with, or oppose.
- Agent selection is salience-ranked: principals and arbiters before stakeholders; amplifiers sparingly; never let a well-covered amplifier crowd out a pivotal principal. KG centrality (
kg_centrality_priors) is a supporting signal, not the ranking.
- Pool size is the main cast, hard-capped at 20 (binding). Every persona must be an entity whose decisions and actions carry forecast signal — no filler "audience" personas by default. A 15-agent simulation of real principals beats an 80-agent simulation padded with generic bystanders: filler agents burn one LLM call per active round each while contributing zero forecast information. Media organizations, journalists, commentators, and analysts are never personas (they are graph context/sources). If the dossier cast exceeds 20, take the top 20 by salience; if fewer, do not pad.
- Depth heuristic: if the persona cannot reason three sentences about the forecast's main uncertainty in character, it is not grounded enough.
- Persona language follows the simulated arena (English personas for US/global-market runs, Chinese for Chinese-audience runs).
Each agent also carries flat behavioral fields: stance (one of supportive / opposing / neutral / observer — relative to the central question), interested_topics (the fault lines it cares about, used for clustering), and an influence tier consistent with the dossier's salience.
3. The world brief (shared reality)
Compose ONE compact world brief (≤ ~1400 chars) injected into EVERY agent's system prompt, so all agents argue inside the same reality instead of hallucinating divergent worlds. Deterministic composition, three blocks in order:
- The central question (first ~400 chars of the forecast question) — what this world is arguing about.
- The situation brief from the dossier: current situation · how it got here (context) · the forces in tension (dynamics) · the 3–6 fault lines actors argue over · the catalysts that would shift things.
- Hot topics — the ≤8 named discussion topics for the run.
Missing blocks shorten the brief; never pad with invented context. The brief is shared ground truth — per-actor beliefs and disagreements live in the personas, not here.
4. Seed posts (ignition from researched stances)
The simulation opens with initial posts that put the researched disagreement on the table:
- Authors: the top-influence / top-tier actors (those with a researched, non-empty stance).
- Content: each author states its researched stance verbatim-faithful on one hot topic — "On {topic} — our position: {stance}". Do not soften, merge, or invent stances.
- Engagement: end roughly half of the seed posts with a discussion hook ("Who gains and who loses if this holds?") so replies have somewhere to go.
- Coverage: seed posts should collectively touch the main fault lines, not five posts on one topic.
- Language matches the persona language of the run.
5. Time, activity & scheduled events
- One round = one simulated hour. Choose the activity profile matching the arena:
china_default — active 08:00–23:00; peak 19–22; off-peak 0–5; morning 6–8; work 9–18.us_business — US business rhythm.global_market — cross-timezone: some cohort is always awake, follow-the-sun peaks.
- Per-agent
active_hours must fit both the profile and the actor (a regulator posts in work hours; a trader around market hours; a media account across peaks).
- Scheduled events from the dated timeline: map the dossier's
key_events onto rounds in [0, total_rounds) preserving order and relative spacing, where total_rounds = min(configured_rounds, execution_round_budget) — never schedule a flashpoint beyond the rounds that will actually run. Attach to each event the most relevant high-influence publisher (the actor whose profile best matches the event) so the event enters the feed as a post from a plausible source.
- Event text is drawn from the researched event description — dated, factual, no invented outcomes.
6. The follow graph (valence-aware, clustered, bridged)
The initial follow structure encodes the researched network so influence flows realistically:
- Relationship-derived follows: for each researched relationship between selected agents, add follow edges consistent with its type and valence — allies/partners follow each other; competitors and opponents follow each other too (adversaries monitor each other); regulators and regulated watch each other. Valence lives in the personas' relations (an opponent-follow is not an endorsement).
- Echo-chamber clustering: cluster agents by
(stance bucket, dominant interested topic) and densify follows within clusters — real discourse has homophily; a sim where everyone follows everyone converges to mush.
- Cross-cluster bridges: high-influence agents keep follow edges ACROSS clusters, so information can jump chambers the way it does through elite accounts in reality.
Never fabricate a relationship edge that the dossier/KG does not support; sparse researched networks stay sparse (the clustering step adds homophily follows, clearly derived from stance/topic, not fake alliances).
7. Output contract (simulation_config.json)
The design lands in the config the simulation engine's sim_start consumes (pass its file path as config_path). Populate:
world_brief — §3 (omit the field entirely if empty; the runner skips injection).agent_configs[] — per agent: canonical name, persona prose, persona_design (§2, under other_info.persona_design), stance, interested_topics, influence tier, active_hours.event_config — initial_posts[] (§4), scheduled_events[] (§5, with round index + publisher), hot_topics[].time_config — profile hours (peak/off-peak/morning/work), rounds ↔ hour mapping.initial_follows[] — §6, as [follower_index, followee_index] pairs.
When run as a pipeline stage, simulation_config.json is the required stage artifact (the driver gates on it); optionally also emit agent_profiles.json (the per-agent profile list) for the manifest.
After sim_start (which returns a sim_id immediately — the run is a detached job), poll sim_status (the run takes 60–120 minutes; never busy-wait in tight loops), then read sim_results; use sim_interview_agents for post-hoc structured probes of specific agents. Do not attempt to run the simulation inside your own turn — it is an external job.
8. Quality gate
9. Failure modes
- ❌ Template personas ("a thoughtful analyst who weighs both sides") — the hollow-sim signature.
- ❌ Inventing a stance, alliance, or red line the dossier does not evidence.
- ❌ Assigning artificial "analytical schools" for diversity instead of using real actor differences.
- ❌ A world brief that editorializes or resolves the fault lines instead of stating them.
- ❌ Seed posts all on one topic, or paraphrased so far they no longer match the researched stance.
- ❌ Scheduling flashpoint events past the execution round budget (they never fire).
- ❌ Fully-connected or valence-blind follow graphs; rival and ally follows treated identically in personas.
- ❌ Amplifier-heavy agent rosters that crowd out the principals who actually move the outcome.
