Brainstorm Swarm

Brainstorm Swarm

v1.0.0 | Multi-perspective brainstorming via parallel persona subagents | 4 skills + 12 personas + 1 command

---

The problem

You're about to make a non-trivial decision — design a feature, pick an architecture, write a strategy doc, change an organizational policy, plan a migration. The right way to think about it is from multiple perspectives: the PM cares about scope and value, the engineer cares about complexity and rollout, the designer cares about user friction, the skeptic asks what could go wrong.

The problem: you don't have a roundtable handy. You're alone with Claude. So you either:

• Ask Claude to "think about this from different perspectives" — and get a single voice approximating multiple roles, badly
• Imagine the perspectives yourself — and miss the ones outside your habits
• Convene a real meeting — and pay the coordination cost for early-stage thinking

The solution

Brainstorm Swarm spawns 6-12 persona-distinct subagents in parallel, each interviewing you from their perspective. Each persona is a real subagent (not Claude playing a role) — so they have isolated context, parallel execution, and genuine multi-perspective output.

After the parallel interview, the orchestrator synthesizes their contributions into:

• Consensus — what every persona agreed on
• Dissent — where personas disagreed and why
• Open questions — what nobody could answer; what to research next
• Recommended next move — the synthesized decision

The 12 personas

Persona	Voice
pm	Product Manager — value, scope, success metrics, prioritization
engineer	Implementation realism — feasibility, complexity, dependencies, rollout
designer	UX — user flow, friction, accessibility, aesthetic coherence
skeptic	Devil's Advocate — what's wrong with this, what assumption hides
user-advocate	The customer's voice in the room — what they actually want
pre-mortem-specialist	Assume failure in 6 months — what killed it
junior	Naive questioner — surfaces buried assumptions, demands explanations
veteran	War stories — "I've seen this before, here's what bit us"
first-principles-thinker	Strip back to fundamentals — what's actually true
constraint-setter	Scope discipline — what's NOT in scope, where does this stop
optimist	Yes-and energy — what's the most ambitious version of this
operator	Production reality — security, monitoring, blast radius, on-call

The 4 skills

Skill	What it teaches
swarm-protocol	The orchestration logic — when to invoke a swarm, which subset of personas, how to spawn them in parallel, how to handle their outputs
interview-facilitation	How to structure the multi-perspective interview as a coherent arc, how to design questions, when to go deep vs broad
swarm-synthesis	How to combine multiple persona outputs into a useful artifact — consensus matrix, dissent log, decision synthesis, what to write down
custom-personas	How to design ad-hoc personas for niche domains when the canonical 12 don't fit (CFO for budget brainstorms, security engineer for threat modeling, etc.)

Installation

/plugin marketplace add viktorbezdek/skillstack
/plugin install brainstorm-swarm@skillstack

Usage

Slash command

/brainstorm-swarm:start adding offline mode to our mobile app

The command activates the swarm-protocol skill, asks 1-2 clarifying questions about scope, proposes a persona subset, gets your sign-off, spawns the chosen subagents in parallel, collects their outputs, and synthesizes.

Auto-activation

The skills also activate on natural-language queries:

• "Brainstorm this with multiple perspectives"
• "Get me a virtual roundtable on this design"
• "Workshop this idea from PM, engineer, designer, and skeptic angles"
• "Pre-mortem this rollout plan"
• "Run a persona swarm for [topic]"
• "I need devil's advocates on this"

Output

A typical run produces a synthesis document with:

## What we brainstormed
[topic, scope, persona subset, why these personas]

## Consensus
- [point all personas agreed on, with attribution to who raised it first]

## Dissent
- PM vs Engineer on [topic]: [the disagreement]

## Open questions
- [questions nobody could answer; what to research]

## Recommended next move
- [synthesized decision with reasoning]

## Per-persona contribution (collapsible)
- PM said: [...]
- Engineer said: [...]
- ...

What this plugin is NOT

• NOT for code review — use code-review (also a swarm, but locked to PR/code analysis)
• NOT for single-perspective interviews — use elicitation (psychological deep-interview techniques) or deep-interview (Socratic single-agent)
• NOT for executing or building things — use team, autopilot, or multi-agent-patterns
• NOT for creating product personas as artifacts — use persona-definition (creates user personas for product/UX research)
• NOT for stakeholder mapping — use persona-mapping (Power-Interest matrices, RACI charts)
• NOT for theoretical multi-agent architecture — use multi-agent-patterns (the patterns themselves, not a runnable swarm)

Conventions

• Each persona subagent has tight scope and a focused system prompt; they don't bleed into each other's territory
• The synthesis skill explicitly preserves dissent — the swarm's value is in the disagreement, not just the consensus
• Custom personas are encouraged when a domain calls for it (the custom-personas skill teaches the design discipline)

Benchmark evidence

Three benchmark rounds (sonnet-4-6 executor + grader, planning evals where the plan must use the canonical 12 persona names, the named arc taxonomy, the single-message parallel-spawn rule, the Phase 3 FOR/AGAINST prompt format, and the brief-vs-fork custom-persona framework):

Round	Format	Δ pass-rate	Verdict
1	skill v1.0.0 (only swarm-protocol loaded)	+0.00	⚪ Indistinguishable — the skill content was buried in references/; the executor invented its own persona names ("Visionary, Pragmatist, Synthesizer") and arc labels ("Pre-Mortem Arc / Red-Team Phase")
2	skill v1.1.0 (canonical names + Arc taxonomy elevated to top of swarm-protocol/SKILL.md)	+0.11	🟡 Weak signal — the targeted fix landed: with-skill plans now use pm, engineer, pre-mortem-specialist, etc. literally. Cross-skill assertions (Phase 3 prompt format, custom-persona template) remained unreachable because their content lives in sibling SKILL.md files the single-skill benchmark didn't load
3	rules v1.1.0 (all 4 SKILL.md files loaded together)	+0.78	🟢 Strong signal — 7 of 9 assertions converted to Signal. Sonnet baseline scored 1/9; with-rules scored 8/9. The plugin's value lands when its skills work together

Honest reading: the plugin is multi-skill by design. The swarm-protocol skill alone teaches orchestration; interview-facilitation teaches the Phase 3 prompt format; swarm-synthesis teaches the artifact structure; custom-personas teaches the brief-vs-fork framework and 4-slot template. When the slash command invokes the workflow, all four skills are in scope — that's the supported usage pattern, and that's where the +0.78 lift is real.

Benchmark snapshots are committed under benchmarks/brainstorm-swarm/ (single-skill rounds) and benchmarks/brainstorm-swarm-rules/ (full-plugin round) for reproduction.

Author

Viktor Bezdek — https://github.com/viktorbezdek

License

MIT.