Crowdcast Multi-Agent Social Simulation
Spawn dozens of AI agents inside Claude Code with a single /crowdcast command to simulate group behavior and predict public reactions — zero dependencies, no server.
The skill operates only within a local .crowdcast/ directory: no network calls, no credential access, minimal scope; it asks the user to confirm configuration after Phase 1 and does not auto-retry failed subagents. Deducted for: no explicit cleanup/rollback path for copied seed files, unconstrained Bash usage, and missing license/source attribution.
Internally consistent: data schemas, per-phase prompts, status verification, resume-from-checkpoint, and error handling (missing simulations, missing seeds, subagent failures) are all covered. This is a static review with no tests or execution evidence, and edge cases (parallel meta. writes, mid-chunk failure states) are only addressed in prose, so the static cap of 10 applies.
Trigger words are explicit and scenarios (forecast/creative simulation) are clear, with concrete command and input formats; but capability boundaries are undeclared (document language, scale limits, no disclaimer that outputs are not real predictions), no Chinese-language support notes, and the entire value depends on the Claude Code Agent tool environment.
Layered structure (orchestrating SKILL.md plus detailed reference prompts plus a data-schema doc) follows progressive disclosure, and version 0.1.0 exists; however no license information in the skill files, no changelog, no known-limitation disclosure, and no stated maintenance/update path; publisher unverified.
The pipeline is complete (analyze→profile→simulate→report) with well-defined output formats, and resume/interview add value; but simulation output is inherently LLM-generated fiction with unverified predictive value, may need substantial human review, and cost (many subagent calls) versus benefit is unproven — static cap 7, scored 6.
Only author-authored architecture notes and example walkthroughs; no test suite, no CI evidence, no third-party execution records; example outputs are fictional demo data, not real run artifacts — static cap 5, scored 3.
- This is a static source review only; nothing was executed and confidence is low.
- Simulation outputs are LLM-generated fiction and must not be treated as real predictions of public reaction or events.
- The skill creates a .crowdcast/ directory and copies seed files with no automatic cleanup; delete manually when no longer needed.
- Skill files carry no license or changelog; the publisher is unverified — review the repository yourself before adoption.
- No Chinese-language adaptation notes; core function depends entirely on the Claude Code Agent tool environment and will not run elsewhere.
What does this skill do, and when should you use it?
Crowdcast is a Claude Code skill that packages multi-agent social simulation into a single slash command. You supply a document and a question; it analyzes the document, generates agent personas with distinct personalities and stances, runs rounds of interaction on a simulated platform, and produces either a forecast report or a narrative retelling. All state lives as JSON files in a local .crowdcast directory, so interrupted runs can resume. Compared to MiroFish, its headline advantage is zero setup: no Docker, databases, external APIs, or monthly fees — everything runs within your Claude subscription.
Triggered by /crowdcast simulate, it runs four phases: (1) Analyze — a subagent extracts entities and relationships from your documents into a knowledge graph; (2) Profile — 2-4 parallel subagents generate personas for key agents (e.g., influencers) and crowd groups; (3) Simulate — simulator subagents are dispatched sequentially in chunks of ~25 rounds, modeling social-media posts (forecast) or free-form narrative interaction (creative); (4) Report — a final subagent writes report.md plus structured report_data.. Supporting commands: /crowdcast analyze (analysis only), resume (checkpoint recovery), report (regenerate report), and interview (chat in-character with a simulated agent in the main context).
- Policy or PR professionals: feed in a policy document or press release and simulate how 50+ stakeholders would react on social media before it ships.
- Fiction writers: give it existing chapters, let characters interact in creative mode to continue the story, then interview each character to check consistency.
- Content and communications teams: run a low-cost dry run of public reaction to a sensitive announcement in forecast mode.
- Researchers and analysts: use /crowdcast analyze to extract a knowledge graph from documents without running a full simulation.
- Anyone running long jobs: if a simulation crashes midway, /crowdcast resume picks up from the last checkpoint instead of restarting.
What are this skill's strengths and limitations?
- Genuinely zero external dependencies: no Docker, database, web server, or paid external APIs — costs are covered by your Claude subscription
- Resumable by design: every phase persists to JSON, so failures don't force a restart from scratch
- Hybrid depth architecture: key agents think individually while crowd agents are batched, balancing quality and efficiency
- Unique interview mode: converse directly with simulated agents in character to sanity-check results or aid creative work
- Deeply tied to Claude Code's subagent orchestration (parallel/sequential Agent dispatch) — not portable as-is to other platforms
- Practical agent cap of ~100, unsuitable for million-scale simulations
- A typical forecast run takes 20-40 minutes across 50-100 rounds
- Version 0.1.0 with no test suite mentioned in the source; README claims MIT but repo metadata lists license as unknown — verify before adopting
- Requires a paid Claude subscription (Pro, Team, or Enterprise)
How do you install this skill?
Any of three options: 1) Clone directly into the skills directory: git clone https://github.com/TheQmaks/crowdcast.git ~/.claude/skills/crowdcast ; 2) Clone then symlink: ln -s $(pwd)/crowdcast ~/.claude/skills/crowdcast ; 3) Install as a Claude Code plugin: /plugin marketplace add TheQmaks/crowdcast, then /plugin install crowdcast@theqmaks. No npm, pip, Docker, or .env files required. Verify by typing /crowdcast in Claude Code — the help menu should appear.
How do you use this skill?
Full forecast simulation: /crowdcast simulate ./news_report.pdf "How will the public react to this policy change?" (add --mode=forecast or --mode=creative to override auto-detection). Creative continuation: /crowdcast simulate ./chapter1.txt "Continue the story with these characters". Analysis only: /crowdcast analyze ./report.pdf. Resume an interrupted run: /crowdcast resume sim_a3f8b2c91d04. Regenerate a report: /crowdcast report sim_a3f8b2c91d04. Interview an agent: /crowdcast interview sim_a3f8b2c91d04 mayor_chen; type "exit interview" to leave.
How does this skill compare with similar options?
The README explicitly benchmarks against MiroFish, a multi-agent prediction engine powered by OASIS: MiroFish requires Python + Node.js + Docker, depends on Zep Cloud ($25/month) plus per-run LLM API costs, and limits its free tier to 1 simulation/month, but scales to millions of agents. Crowdcast runs from a single git clone with no external fees but caps at ~100 agents. The README's own verdict: choose MiroFish for massive scale, Crowdcast for zero-setup quick predictions.