AgentSkillsCN

think

运用平行智能体进行深度结构化思考(内部)

SKILL.md
--- frontmatter
name: think
description: Deep structured thinking with parallel agents (INTERNAL)
tools: Read, Write, Bash, Task

/think - Deep Structured Thinking

INTERNAL SKILL — Used by @idea and @design for deep analysis.

When to Use

  • Complex tradeoffs with no clear answer
  • Architectural decisions with multiple valid approaches
  • Unknown unknowns in requirements
  • System-level implications

Parallel Expert Agents Pattern

Step 1: Define Expert Roles

ExpertFocusWhen to Use
ArchitectSystem design, patternsAll architectural decisions
SecurityThreats, auth, dataUser data, APIs, external integration
PerformanceLatency, scalabilityHigh load, real-time
UXUser experienceUser-facing features
OpsDeployability, monitoringProduction systems

Step 2: Launch Parallel Analysis

python
# Spawn 2-4 experts in parallel (single message)
Task(
    subagent_type="general-purpose",
    prompt="""You are the ARCHITECT expert.
    
PROBLEM: {problem}

Analyze from your perspective:
1. Key considerations?
2. Applicable patterns?
3. Risks?

Return 3-5 bullet points.""",
    description="Architect analysis"
)
# Launch other experts similarly...

Step 3: Synthesize

After all experts complete:

markdown
## Expert Analysis

**@architect:** Domain layer first, risk of tight coupling
**@security:** OAuth2 preferred, need rate limiting
**@performance:** Caching needed, ~500MB for 10K users

## Synthesis
Recommended approach combining all perspectives...

## Open Questions
What remains unknown...

Single-Agent Mode (Simple Problems)

  1. Deconstruct problem into dimensions
  2. Explore 3+ angles (ideal/pragmatic/minimal)
  3. Synthesize insights
  4. Present findings with tradeoffs

Output Format

markdown
## Problem Analysis

### Context
{Brief problem statement}

### Expert Analysis
**@architect:** {analysis}
**@security:** {analysis}

### Synthesis
{Combined insights}

### Recommendation
{Clear recommendation with rationale}

### Open Questions
{What remains unknown}

Principles

  • Parallel exploration — Multiple experts simultaneously
  • Role-based expertise — Each expert has defined perspective
  • Explicit tradeoffs — State what you're optimizing for
  • Clear recommendation — Don't leave user hanging