AgentSkillsCN

ai-conversation

以持久化上下文与审计轨迹,统筹多AI对话。借助“邪恶看板”作为共享记忆,让Claude、Gemini、Codex与OpenCode能够协同合作。 适用场景如下: - 对架构或设计进行多模型分析 - 就某一决策获取多元化的AI视角 - 需要可审计的AI辅助决策 - 从多源AI中凝聚共识

SKILL.md
--- frontmatter
name: ai-conversation
description: |
  Orchestrate multi-AI conversations with persistent context and audit trails.
  Use wicked-kanban as shared memory so Claude, Gemini, Codex, and OpenCode can collaborate.

  Use when:
  - Running multi-model analysis on architecture or design
  - Getting diverse AI perspectives on a decision
  - Need auditable AI-assisted decision making
  - Building consensus from multiple AI sources

Multi-AI Conversation Orchestration

Purpose: Coordinate conversations across multiple AI models with shared context, persistent state, and audit trails.

Why This Matters

Single-AI analysis has blind spots. Multi-model conversations:

  • Catch more issues: Different models flag different concerns
  • Build confidence: Consensus across models = higher confidence
  • Create audit trails: Document which AI said what, when
  • Persist context: Insights survive session restarts

Most users query one AI. This skill teaches systematic multi-model orchestration.

Quick Usage

Inline (Simple)

markdown
**Task**: Review auth design

**Multi-model check**:
- Claude: [your analysis in this conversation]
- Gemini: `cat design.md | gemini "Review for security"`
- Codex: `cat design.md | codex exec "Review for architecture"`

**Consensus**: All flagged token expiry. High confidence issue.

Structured (Complex)

For important decisions, use full orchestration:

  1. Create kanban task for shared context
  2. Gather perspectives systematically
  3. Synthesize with documented rationale
  4. Store decision in wicked-mem

The Orchestration Process

Step 1: Establish Shared Context

Create a single source of truth all AIs can reference.

bash
/wicked-kanban:new-task "Design review: Payment API" --priority P0

Add the context document to the task description.

→ See refs/context.md for context management patterns.

Step 2: Gather Perspectives

Query each AI with the same focused prompt for comparable results.

AICommandStrength
Claude(in conversation)Nuanced reasoning
Geminicat doc | gemini "prompt"Long context
Codexcat doc | codex exec "prompt"Code-focused
OpenCodeopencode run "prompt" -f docMulti-provider

Add each response as a kanban comment with attribution.

→ See refs/automation.md for scripts.

Step 3: Synthesize Results

Compare perspectives and identify:

SignalMeaningAction
Consensus (2+ agree)High confidence issueAddress immediately
Unique insightOne AI caught somethingEvaluate carefully
DisagreementGenuine tradeoffHuman decides
SilenceNo AI flagged itLower priority

→ See refs/examples.md for synthesis templates.

Step 4: Record and Persist

Store the decision with full attribution:

bash
/wicked-mem:store "Payment API: Use Stripe webhooks.
Consensus: Claude, Gemini, Codex (idempotency critical).
Unique: OpenCode flagged circuit breaker need.
Decision by: [Human] on [Date]" --type decision --tags payments

→ See refs/auditability.md for audit patterns.

Decision Principles

When to Use Multi-Model

SituationRecommendation
Architecture decisionsYes - high impact
Security reviewYes - catch blind spots
Quick bug fixNo - overhead not worth it
Important PRsYes - diverse review
Routine codeNo - single AI sufficient

How to Weight Perspectives

  1. Consensus wins: 2+ models agreeing = high confidence
  2. Expertise matters: Codex for code, Gemini for long docs
  3. Unique insights valuable: Don't dismiss outliers
  4. Human decides conflicts: AI informs, human chooses

When to Start Fresh vs. Continue

SignalAction
New topicFresh context
Building on priorContinue session
Context pollutedSummarize and restart
Need unbiased viewFresh, neutral handoff

→ See refs/context.md for session patterns.

Output Format

For formal multi-model reviews:

markdown
## Multi-Model Review: [Topic]

**Context**: [What was reviewed]
**Models**: Claude, Gemini, Codex, OpenCode

### Consensus (High Confidence)
- Issue 1: [flagged by Claude, Gemini, Codex]
- Issue 2: [flagged by all]

### Unique Insights
- **Gemini**: [insight others missed]
- **OpenCode**: [insight others missed]

### Disagreements
- [Topic]: Claude says X, Codex says Y

### Decision
[What was decided and why]

### Stored
- wicked-mem: [memory ID]
- wicked-kanban: [task ID]

Quick Reference

NeedSolution
Shared contextwicked-kanban task
Persist decisionswicked-mem
Automate gatheringrefs/automation.md scripts
Audit trailrefs/auditability.md patterns
Session managementrefs/context.md
Templatesrefs/examples.md

References

Process:

Quality:

  • Auditability - Audit trails, compliance, decision tracking
  • Examples - ADR templates, synthesis patterns