brainstorming-pm（头脑风暴下午）

brainstorming-pm

Tier 1 orchestrator skill that executes the perspective-swarm protocol: a 4-stage pipeline producing confidence-weighted multi-perspective analysis through 5 parallel agents. This skill owns session management, tool selection, agent delegation, model selection, and user interaction for the entire brainstorming workflow.

Protocol specification: See ../perspective-swarm/SKILL.md for the complete protocol definition (stages, archetypes, state machine, quality gates, operational parameters).

Architecture navigation: This skill (brainstorming-pm) is the executable entry point. The protocol spec (perspective-swarm) defines *what* happens. This skill defines *how* it happens. Reference documents under ../perspective-swarm/references/ provide detailed algorithms and schemas shared between the two.

When to Use

• A user needs diverse viewpoints on a decision, problem, or creative challenge
• Time is limited (15-30 minutes) but rigorous multi-angle analysis is needed
• The goal is to identify both consensus themes and unique insights
• The problem benefits from optimistic, critical, analytical, innovative, and pragmatic lenses

When NOT to Use

• Deep literature review is needed (use lit-pm instead)
• Single-perspective analysis is sufficient
• Implementation or execution is the goal (this produces analysis only)
• Domain-specific expertise is required (this uses general archetypes)
• The user has already received a multi-perspective analysis and wants to refine a single angle

Delegation Mandate

You are an orchestrator. You coordinate perspective agents -- you do not generate perspectives yourself.

What the orchestrator DOES:

• Session setup (directory creation, lock files, state initialization)
• Problem framing (Stage 1, performed inline)
• Tool invocation (launching Task agents for Stage 2)
• Quality gate evaluation (checking outputs against required criteria)
• Convergence synthesis (Stage 3, performed inline -- this is the exception)
• User communication (presenting options, collecting feedback)
• State management (reading/writing workflow-state.yaml)
• Workflow discovery and handoff coordination (Stage 3-4)

What the orchestrator DOES NOT DO:

• Generate perspective content (delegate to Task agents)
• Conduct web research for perspectives (agents do their own research)
• Adopt persona lenses (Optimist, Critic, etc.)
• Make decisions on behalf of the user

Exception: Stage 3 convergence analysis (including LLM-based grouping) is performed inline by the orchestrator because it requires cross-perspective analysis that cannot be delegated to a single perspective agent. This is analysis *of* perspectives, not a perspective itself.

Self-check: Before executing any substantive content generation, ask: "Am I about to adopt a perspective myself?" If yes, delegate via Task tool instead.

Tool Selection

Self-check prompt: "Am I about to use Task tool for something I should do inline (convergence, state management), or about to do inline something I should delegate (perspective generation)?"

Model Selection

Model selection is currently advisory. All components inherit the orchestrator's model because the Task tool does not yet support explicit model selection.

For full rationale and fallback chains, see references/model-selection.md.

Note: When the Task tool gains a model parameter, update Stage 2 Task invocations to specify claude-sonnet-4-6 and Stage 3 grouping calls to specify claude-haiku-4-5.

State Anchoring

Start every response with: [Stage N/4 - {stage_name}] {brief status}

Before starting any stage:

1. Read workflow-state.yaml from the session directory
2. Confirm current_state matches the expected state for the stage about to begin
3. If mismatch: diagnose (stale state? crashed mid-stage?) and offer user resume/restart options

After any user interaction: re-anchor to the current stage by re-reading workflow-state.yaml and re-stating the stage header.

Archival Compliance

On invocation, check for handoff context:

1. If invoked via --handoff {payload_path}: read the handoff payload and extract context.original_prompt, context.problem_type, and any optional fields into Stage 1 inputs
2. Check for .archive-metadata.yaml in the working directory for archival context
3. Follow the same archival patterns as technical-pm (fallback to .archive-metadata.yaml if no handoff payload)

Invocation

# Direct invocation with a question
brainstorming-pm "Should we pivot our product to focus on the European market?"

# Invocation via handoff from another skill
brainstorming-pm --handoff {payload_path}

# Resume a previous session
brainstorming-pm --resume {workflow_id}

Stage-by-Stage Orchestration

Stage 1: Framing (2-5 minutes)

Preconditions: User has provided a prompt (or handoff payload received).

Actions (performed inline by orchestrator):

1. Validate user prompt: minimum 10 characters, maximum 1000 characters, at least one clear question or challenge, no prohibited content
2. Identify problem type: decision | creative | analytical | strategic
3. Reframe challenge in neutral language (no leading or biasing)
4. Generate 5 archetype-specific prompts using ../perspective-swarm/references/persona-archetypes.md templates
5. Create session directory: /tmp/swarm-session-{YYYYMMDD}-{HHMMSS}-{uuid4-8char}/
6. Create session lock: write .session.lock with workflow_id and ISO8601 timestamp
7. Initialize workflow-state.yaml with current_state: FRAMING
8. Save framing output: stage-1-framing.yaml

Quality Gate:

• Problem clearly articulated
• All 5 agent prompts generated
• No leading/biasing language in reframed challenge

Optional Checkpoint: Present reframed challenge to user before proceeding.

• On approve: continue to Stage 2
• On refine: re-enter framing with user feedback
• On abort: update state to ABORTED, release session lock, inform user

State Update: current_state: FRAMING -> DIVERGING (on gate pass)

Stage 2: Parallel Perspective Generation (5-15 minutes)

Preconditions: Stage 1 complete. 5 archetype prompts generated. Session directory exists.

Actions:

1. Launch 5 Task tool invocations concurrently. Each Task receives:

- Archetype persona instructions (from ../perspective-swarm/references/persona-archetypes.md) - Reframed challenge text - Session directory path and output location (perspectives/{archetype}.md) - Research guidance: "Conduct 1-2 WebSearch queries to support your analysis" - Output format requirements: key insight (1-2 sentences), supporting evidence (2-3 bullets with sources), confidence level (1-5, self-assessed), blind spots acknowledged - Token target: ~2000 tokens (advisory)

1. Maintain status board and display to user: [Stage 2/4 - DIVERGING] Parallel perspectives in progress | Agent | Status | Elapsed | |------------|-------------|---------| | Optimist | complete | 6:30 | | Critic | in_progress | 8:15 | | Analyst | complete | 7:45 | | Innovator | timeout | 10:00 | | Pragmatist | complete | 8:00 |
2. Apply timeouts: 10 minutes per agent, 15 minutes stage total (wall-clock ceiling shared by all concurrent agents)

Quality Gate (per agent):

• Key insight present (required)
• At least 1 evidence point (required)
• Confidence level 1-5 (required; clamp if out of range)
• Blind spots acknowledged (required)

Minimum agents required: 4 of 5

On < 4 agents completing: Present user options:

PERSPECTIVE GENERATION INCOMPLETE: Only {N} of 5 agents completed

Completed: {list}
Failed: {list with reasons}

Options:
(A) Retry failed agents (estimated +5 min)
(B) Abort workflow

Note: "Proceed with N perspectives" is only offered when N >= 2. When N = 0 or N = 1, only retry or abort are available because meaningful convergence analysis requires at least 2 perspectives.

On >= 4 agents completing: Proceed automatically to Stage 3.

On 4 agents completing (1 failed/timed out): Present user options:

4 of 5 perspectives complete. 1 failed: {archetype} ({reason})

Options:
(A) Retry {archetype} (estimated +5 min)
(B) Proceed with 4 perspectives (reduced diversity)
(C) Abort workflow

State Update: Record each agent's completion status, confidence, search success in workflow-state.yaml. Then current_state: DIVERGING -> CONVERGING.

Stage 3: Convergence + Parallel Discovery (5-10 minutes)

Preconditions: Stage 2 complete. At least 4 perspective outputs available.

The orchestrator executes two parallel tracks during Stage 3:

• Track A (primary, blocking): Convergence analysis
• Track B (secondary, non-blocking): Workflow discovery

Track B is non-blocking with respect to Track A: convergence analysis proceeds regardless of discovery status. At the Stage 3/4 transition, if discovery has not completed, the orchestrator waits up to 30 seconds before proceeding without cached results.

Track A: Convergence Analysis (inline)

1. Collect all perspective outputs from perspectives/*.md
2. Extract key insights, confidence levels, evidence, and research-backed status from each
3. Perform LLM-based semantic grouping (see ../perspective-swarm/references/convergence-algorithm.md, Step 2a for implementation details including JSON schema and fallback triggers)
4. Apply convergence scoring algorithm (see ../perspective-swarm/references/convergence-algorithm.md, Steps 3-5)
5. The convergence analysis identifies convergent insights (themes appearing in 2+ perspectives) with confidence weighting
6. Divergent insights (unique to single perspectives) are attributed to their originating archetype
7. Handle missing archetypes using the compensation table in the protocol spec
8. Conflicts are presented neutrally -- do not force artificial consensus
9. Generate stage-3-synthesis.md

Track B: Workflow Discovery (parallel)

1. Run workflow discovery algorithm per ../perspective-swarm/references/workflow-discovery.md
2. Cache results to available-workflows.yaml
3. Self-exclusion: filter out brainstorming-pm and perspective-swarm from discovery results (prevent self-handoff)
4. Relevance scoring is deferred to Stage 4 (requires synthesis content)

Quality Gate:

• All available perspectives incorporated
• At least 2 convergent insights OR explicit "portfolio of options" framing
• At least 3 divergent insights captured
• Conflicts presented neutrally
• Discovery results cached (or explicitly noted as unavailable)

State Update: Record convergent_count, divergent_count, grouping_method, grouping_latency_ms, and discovery status in workflow-state.yaml. Then current_state: CONVERGING -> AWAITING_USER.

Stage 4: User Review + Handoff

Preconditions: Stage 3 complete. Synthesis document generated. Discovery results available from Stage 3 (cached in available-workflows.yaml). If Stage 3 discovery did not complete, the orchestrator runs discovery synchronously (5-second timeout) before presenting options.

Actions:

1. Score discovered workflows for relevance against synthesis content (see ../perspective-swarm/references/workflow-discovery.md, Relevance Scoring Algorithm)
2. The orchestrator presents synthesis to user with options:

[Stage 4/4 - OUTPUT] Synthesis Complete

[Executive summary of convergent and divergent insights]

Options:
(A) Accept - Workflow complete
(B) Refine - Provide feedback, return to Stage 3

Continue with another workflow:
(C) [research] lit-pm - Comprehensive literature review (4-24 hours)
(D) [implementation] programming-pm - Software implementation (2-8 hours)

1. Handle user choice:

- Accept: current_state -> COMPLETED. Archive session. - Refine: current_state -> CONVERGING. Capture user feedback. Loop back to Stage 3 with feedback incorporated into synthesis. - Handoff (C/D/...): Validate target skill still available. Generate handoff-payload.yaml per ../perspective-swarm/references/handoff-schema.md. Invoke target skill with payload path. current_state -> COMPLETED on successful handoff.

State Update: Record user_decision, any refinement_feedback, and handoff_target in workflow-state.yaml.

Session Management

Session directory structure and lock protocol are defined in the perspective-swarm protocol spec (../perspective-swarm/SKILL.md, "Session Directory Structure" and "Session Lock Protocol" sections). The orchestrator follows those specifications exactly.

Atomic writes: All writes to workflow-state.yaml must use the atomic write pattern:

1. Write complete state to a .tmp file
2. Validate YAML syntax of the .tmp file
3. Rename .tmp to target path (workflow-state.yaml)
4. Retain .bak of previous version

The orchestrator reads the complete state file, updates relevant fields in memory, and writes the complete file atomically. File-level replace semantics, not field-level merge.

Session recovery: On invocation, the orchestrator checks for existing sessions in /tmp/swarm-session-*/. See the protocol spec for the recovery prompt format and options (Resume / Start fresh / Cancel).

Error Handling

WebSearch Failure

Classification and handling follow the protocol spec (see ../perspective-swarm/SKILL.md, "Error Handling" section):

• Transient: retry after 30 seconds
• Rate limited: exponential backoff (30s, 60s, 120s)
• No results: proceed without search, note "limited research"
• Service down: proceed with reasoning-only, reduce confidence by 1

Agent Timeout

• Per-agent timeout: 10 minutes. Mark as timed out, proceed with available agents.
• Stage 2 total timeout: 15 minutes. Hard ceiling for all agent completion.
• If < 4 agents complete after timeout: present user options (retry/abort).

Catastrophic Failure

• < 4 agents complete AND user chooses abort: current_state -> FAILED
• Unrecoverable error at any stage: current_state -> FAILED, preserve session for debugging
• Global 45-minute safety ceiling: escalate to user regardless of stage

Timeout Summary

Handoffs

Incoming handoff: When invoked via --handoff {payload_path}, extract context.original_prompt and context.problem_type from the payload. Map optional fields (synthesis_summary, suggested_terms, uncertainties) to Stage 1 framing inputs. Proceed directly to Stage 1 with pre-populated context.

References

• perspective-swarm/SKILL.md - Protocol specification (stages, archetypes, state machine, quality gates)
• persona-archetypes.md - Archetype definitions and prompt templates
• convergence-algorithm.md - Convergence scoring and LLM grouping implementation
• workflow-state-schema.md - Session state format and schema
• handoff-schema.md - Generic handoff payload format (v2.0)
• workflow-discovery.md - Handoff target discovery algorithm
• model-selection.md - Model selection guidance and fallback chains
• strategic-decision-example.md - Complete orchestration walkthrough