ischca/swarm-discussion-skill
Overview
This skill runs multi-agent, team-based discussions that replicate Staff+ engineer thinking for unsolved or high-uncertainty problems. It composes dynamic and fixed expert personas, enforces structured disagreement, and preserves an argument graph so you get traceable, debatable reasoning rather than single-point answers. Use it to expose blind spots, test assumptions, and surface decision tradeoffs.
How this skill works
The orchestrator generates a team of experts (dynamic + fixed roles like Moderator, Historian, Contrarian, Cross-Domain) and a tension map that forces opposing viewpoints. Discussion proceeds in phased rounds: position declarations, steel-manning, targeted rebuttals, and convergence checks, with live progress reports and automatic quality gates. All messages, position shifts, and citation chains are saved to produce an argument graph and reproducible artifacts.
When to use it
- Unprecedented or high-stakes design decisions
- Complex tradeoffs requiring multiple domain perspectives
- Situations with high risk of groupthink or premature consensus
- When you need auditable reasoning and traceable evidence
- Exploratory research where multiple hypotheses should be stress-tested
Best practices
- Define a clear problem statement and success criteria before spawning the team
- Choose the discussion mode (lightweight/standard/deep) to match budget and complexity
- Review and confirm generated experts and the tension map before starting
- Require wouldChangeIf and confidence fields in position declarations to prevent unfalsifiable claims
- Use the Moderator’s prompts and disagreement budget to keep debates productive and bounded
Example use cases
- Choosing an architectural pattern for distributed transactions in a microservices platform
- Evaluating safety and deployment strategy for a novel ML model in production
- Comparing vendor solutions with cross-domain tradeoffs (cost, security, scalability)
- Stress-testing incident response plans by having Contrarian and Cross-Domain play devil’s advocate
- Running rapid sanity checks on product design options with a lightweight round
FAQ
Mode controls duration and calls: lightweight is 1–2 rounds, standard 2–3, deep 3–5. More experts, rounds, and evidence requests increase compute and cost; pick a mode that balances thoroughness and budget.
Can I modify experts or tensions mid-discussion?
Yes. The Moderator prompts for confirmations and allows modifications after initialization. Position shifts are recorded by the Historian so changes remain auditable.