swarm-meeting（群体会议）

Use this skill to turn a vague "let's discuss" request into a bounded decision meeting.

Inputs

Provide:

• topic
• need
• profile name
• optional artifacts, constraints, decision owner, and timebox

If the request spans many files or docs, prepare a brief pack first with swarm-research-pack or by reading the smallest relevant local artifacts directly.

Ambiguous Input Assist

When the user asks for a meeting without enough structure, do not stop immediately. Convert the vague request into a short guided start.

1. Infer balanced as the default profile unless the user clearly asks for stronger conflict or a deeper architecture review.
2. Infer a default need of options, objections, and recommendation when the user does not specify one.
3. Rephrase the likely topic from the active conversation or nearest artifact.
4. Show a short kickoff block before the meeting if key inputs are missing.

Use a compact pattern like this:

I can run this as a `balanced` swarm meeting.

- Topic: <inferred topic or "please provide one sentence">
- Need: options, objections, and recommendation
- Other profiles: `architecture-debate`, `extreme-diversity`

Examples:
- `Use the swarm-meeting skill. Topic: ... Need: ...`
- `Use swarm-meeting with profile: architecture-debate ...`

Keep this assist block short. Do not turn it into a questionnaire unless the missing information would materially change the outcome.

Workflow

1. Load exactly one profile from profiles/.
2. Load only the role prompts needed for that profile from prompts/.
3. Rephrase the decision in 1-2 sentences and define the agenda.
4. Collect independent role memos before synthesis. Keep each memo short: 3-7 bullets or short paragraphs.
5. Run the objection round if the profile requires it.
6. Force a recommendation, open questions, and concrete next actions.
7. Separate facts from inferences. Mark assumptions explicitly.

Guardrails

• Treat roles as evaluation biases, not personalities.
• Keep max_agents between 4 and 6 unless the user explicitly asks for a larger meeting.
• Prefer local role emulation by default. Only spawn subagents when the user explicitly asks for delegation or parallel agents, or when the environment clearly allows it.
• Do not let any single role dominate the summary.
• Avoid unbounded brainstorming. End with a decision or a clearly framed fork.
• If the user input is underspecified, provide the short assist block above, make one reasonable default assumption, and continue.

Output Contract

Use this section order unless the user asks for a different format:

# Design Review
## Problem Statement
## Agenda
## Role Memos
## Options Considered
## Recommended Architecture
## MVP Plan
## Risks and Failure Modes
## Open Questions
## Concrete Next Actions

Resources

• profiles/balanced.yaml: default 5-role review for most implementation decisions.
• profiles/architecture-debate.yaml: stronger architecture pressure with semantics and platform concerns.
• profiles/extreme-diversity.yaml: intentionally conflicting evaluation axes for high-stakes tradeoffs.
• prompts/*.md: short role cards. Load only selected roles.
• references/prompt-examples.md: ready-to-copy invocation examples for common meeting intents.
• references/operating-rules.md: explicit rules for when to stay local versus when to spawn subagents.
• references/task-scoped-context-review.md: migrated example from the original one-off prompt.
• scripts/select_profile.py: recommend or inspect a profile without external dependencies.
• scripts/render_meeting_context.py: render a meeting brief from a profile plus topic, need, and artifacts.

Role Selection Rules

• Always include moderator.
• Include skeptic unless the user explicitly wants a fast, low-friction pass.
• Prefer pragmatist when delivery pressure is real.
• Prefer minimalist for MVP, operational simplicity, or "ship this week".
• Prefer maximalist for platform or long-horizon design.
• Prefer retrieval-architect when the topic involves search, indexing, code selection, Tree-sitter, LSP, or semantics.
• Prefer granary-workflow-lead when the topic involves task hubs, handoff, checkpointing, CLI flow, or session lifecycle.
• Prefer semantic-purist when correctness depends on type, symbol, or compiler information.
• Prefer lightweight-contrarian when a simple rg/diff/docs index may beat a heavy platform.
• Prefer context-economist when token budget, context pollution, or fresh-session strategy matters.
• Prefer workflow-conservative when tool sprawl is the main risk.