mirror of
https://github.com/catlog22/Claude-Code-Workflow.git
synced 2026-02-05 01:50:27 +08:00
Add discuss-plan command for multi-model collaborative planning and analysis
- Introduced the discuss-plan command to facilitate iterative discussions among Gemini, Codex, and Claude. - Defined roles and priorities for each model in the discussion loop. - Established a structured workflow for input processing, discussion rounds, and user review. - Implemented TodoWrite tracking for progress across rounds. - Specified output routing for logs and final plans. - Included examples and best practices for effective usage.
This commit is contained in:
catlog22
308
.claude/commands/cli/discuss-plan.md
Normal file
308
.claude/commands/cli/discuss-plan.md
Normal file
@@ -0,0 +1,308 @@
|
||||
---
|
||||
name: discuss-plan
|
||||
description: Orchestrates an iterative, multi-model discussion for planning and analysis without implementation.
|
||||
argument-hint: "[--topic '...'] [--task-id '...'] [--rounds N]"
|
||||
allowed-tools: SlashCommand(*), Bash(*), TodoWrite(*), Read(*), Glob(*)
|
||||
---
|
||||
|
||||
# CLI Discuss-Plan Command (/cli:discuss-plan)
|
||||
|
||||
## Purpose
|
||||
|
||||
Orchestrates a multi-model collaborative discussion for in-depth planning and problem analysis. This command facilitates an iterative dialogue between Gemini, Codex, and Claude (the orchestrating AI) to explore a topic from multiple perspectives, refine ideas, and build a robust plan.
|
||||
|
||||
**This command is for discussion and planning ONLY. It does NOT modify any code.**
|
||||
|
||||
## Core Workflow: The Discussion Loop
|
||||
|
||||
The command operates in iterative rounds, allowing the plan to evolve with each cycle. The user can choose to continue for more rounds or conclude when consensus is reached.
|
||||
|
||||
```
|
||||
Topic Input → [Round 1: Gemini → Codex → Claude] → [User Review] →
|
||||
[Round 2: Gemini → Codex → Claude] → ... → Final Plan
|
||||
```
|
||||
|
||||
### Model Roles & Priority
|
||||
|
||||
**Priority Order**: Gemini > Codex > Claude
|
||||
|
||||
1. **Gemini (The Analyst)** - Priority 1
|
||||
- Kicks off each round with deep analysis
|
||||
- Provides foundational ideas and draft plans
|
||||
- Analyzes current context or previous synthesis
|
||||
|
||||
2. **Codex (The Architect/Critic)** - Priority 2
|
||||
- Reviews Gemini's output critically
|
||||
- Uses deep reasoning for technical trade-offs
|
||||
- Proposes alternative strategies
|
||||
- **Participates purely in conversational/reasoning capacity**
|
||||
- Uses resume mechanism to maintain discussion context
|
||||
|
||||
3. **Claude (The Synthesizer/Moderator)** - Priority 3
|
||||
- Synthesizes discussion from Gemini and Codex
|
||||
- Highlights agreements and contentions
|
||||
- Structures refined plan
|
||||
- Poses key questions for next round
|
||||
|
||||
## Parameters
|
||||
|
||||
- `<input>` (Required): Topic description or task ID (e.g., "Design a new caching layer" or `PLAN-002`)
|
||||
- `--rounds <N>` (Optional): Maximum number of discussion rounds (default: prompts after each round)
|
||||
- `--task-id <id>` (Optional): Associates discussion with workflow task ID
|
||||
- `--topic <description>` (Optional): High-level topic for discussion
|
||||
|
||||
## Execution Flow
|
||||
|
||||
### Phase 1: Initial Setup
|
||||
|
||||
1. **Input Processing**: Parse topic or task ID
|
||||
2. **Context Gathering**: Identify relevant files based on topic
|
||||
|
||||
### Phase 2: Discussion Round
|
||||
|
||||
Each round consists of three sequential steps, tracked via `TodoWrite`.
|
||||
|
||||
**Step 1: Gemini's Analysis (Priority 1)**
|
||||
|
||||
Gemini analyzes the topic and proposes preliminary plan.
|
||||
|
||||
```bash
|
||||
# Round 1: CONTEXT_INPUT is the initial topic
|
||||
# Subsequent rounds: CONTEXT_INPUT is the synthesis from previous round
|
||||
~/.claude/scripts/gemini-wrapper -p "
|
||||
PURPOSE: Analyze and propose a plan for '[topic]'
|
||||
TASK: Provide initial analysis, identify key modules, and draft implementation plan
|
||||
MODE: analysis
|
||||
CONTEXT: @{CLAUDE.md} [auto-detected files]
|
||||
INPUT: [CONTEXT_INPUT]
|
||||
EXPECTED: Structured analysis and draft plan for discussion
|
||||
RULES: Focus on technical depth and practical considerations
|
||||
"
|
||||
```
|
||||
|
||||
**Step 2: Codex's Critique (Priority 2)**
|
||||
|
||||
Codex reviews Gemini's output using conversational reasoning. Uses `resume --last` to maintain context across rounds.
|
||||
|
||||
```bash
|
||||
# First round (new session)
|
||||
codex chat "
|
||||
PURPOSE: Critically review technical plan
|
||||
TASK: Review the provided plan, identify weaknesses, suggest alternatives, reason about trade-offs
|
||||
MODE: analysis
|
||||
CONTEXT: @{CLAUDE.md} [relevant files]
|
||||
INPUT_PLAN: [Output from Gemini's analysis]
|
||||
EXPECTED: Critical review with alternative ideas and risk analysis
|
||||
RULES: Focus on architectural soundness and implementation feasibility
|
||||
"
|
||||
|
||||
# Subsequent rounds (resume discussion)
|
||||
codex chat "
|
||||
PURPOSE: Re-evaluate plan based on latest synthesis
|
||||
TASK: Review updated plan and discussion points, provide further critique or refined ideas
|
||||
MODE: analysis
|
||||
CONTEXT: Previous discussion context (maintained via resume)
|
||||
INPUT_PLAN: [Output from Gemini's analysis for current round]
|
||||
EXPECTED: Updated critique building on previous discussion
|
||||
RULES: Build on previous insights, avoid repeating points
|
||||
" resume --last
|
||||
```
|
||||
|
||||
**Step 3: Claude's Synthesis (Priority 3)**
|
||||
|
||||
Claude (orchestrating AI) synthesizes both outputs:
|
||||
|
||||
- Summarizes Gemini's proposal and Codex's critique
|
||||
- Highlights agreements and disagreements
|
||||
- Structures consolidated plan
|
||||
- Presents open questions for next round
|
||||
- This synthesis becomes input for next round
|
||||
|
||||
### Phase 3: User Review and Iteration
|
||||
|
||||
1. **Present Synthesis**: Show synthesized plan and key discussion points
|
||||
2. **Continue or Conclude**: Prompt user:
|
||||
- **(1)** Start another round of discussion
|
||||
- **(2)** Conclude and finalize the plan
|
||||
3. **Loop or Finalize**:
|
||||
- Continue → New round with Gemini analyzing latest synthesis
|
||||
- Conclude → Save final synthesized document
|
||||
|
||||
## TodoWrite Tracking
|
||||
|
||||
Progress tracked for each round and model.
|
||||
|
||||
```javascript
|
||||
// Example for 2-round discussion
|
||||
TodoWrite({
|
||||
todos: [
|
||||
// Round 1
|
||||
{ content: "[Round 1] Gemini: Analyzing topic", status: "completed", activeForm: "Analyzing with Gemini" },
|
||||
{ content: "[Round 1] Codex: Critiquing plan", status: "completed", activeForm: "Critiquing with Codex" },
|
||||
{ content: "[Round 1] Claude: Synthesizing discussion", status: "completed", activeForm: "Synthesizing discussion" },
|
||||
{ content: "[User Action] Review Round 1 and decide next step", status: "in_progress", activeForm: "Awaiting user decision" },
|
||||
|
||||
// Round 2
|
||||
{ content: "[Round 2] Gemini: Analyzing refined plan", status: "pending", activeForm: "Analyzing refined plan" },
|
||||
{ content: "[Round 2] Codex: Re-evaluating plan [resume]", status: "pending", activeForm: "Re-evaluating with Codex" },
|
||||
{ content: "[Round 2] Claude: Finalizing plan", status: "pending", activeForm: "Finalizing plan" },
|
||||
{ content: "Discussion complete - Final plan generated", status: "pending", activeForm: "Generating final document" }
|
||||
]
|
||||
})
|
||||
```
|
||||
|
||||
## Output Routing
|
||||
|
||||
- **Primary Log**: Entire multi-round discussion logged to single file:
|
||||
- `.workflow/WFS-[id]/.chat/discuss-plan-[topic]-[timestamp].md`
|
||||
- **Final Plan**: Clean final version saved upon conclusion:
|
||||
- `.workflow/WFS-[id]/.summaries/plan-[topic].md`
|
||||
- **Scratchpad**: If no session active:
|
||||
- `.workflow/.scratchpad/discuss-plan-[topic]-[timestamp].md`
|
||||
|
||||
## Discussion Structure
|
||||
|
||||
Each round's output is structured as:
|
||||
|
||||
```markdown
|
||||
## Round N: [Topic]
|
||||
|
||||
### Gemini's Analysis (Priority 1)
|
||||
[Gemini's full analysis and proposal]
|
||||
|
||||
### Codex's Critique (Priority 2)
|
||||
[Codex's critical review and alternatives]
|
||||
|
||||
### Claude's Synthesis (Priority 3)
|
||||
**Points of Agreement:**
|
||||
- [Agreement 1]
|
||||
- [Agreement 2]
|
||||
|
||||
**Points of Contention:**
|
||||
- [Issue 1]: Gemini suggests X, Codex suggests Y
|
||||
- [Issue 2]: Trade-off between A and B
|
||||
|
||||
**Consolidated Plan:**
|
||||
[Structured plan incorporating both perspectives]
|
||||
|
||||
**Open Questions for Next Round:**
|
||||
1. [Question 1]
|
||||
2. [Question 2]
|
||||
```
|
||||
|
||||
## Examples
|
||||
|
||||
### Example 1: Multi-Round Architecture Discussion
|
||||
|
||||
**Command**: `/cli:discuss-plan --topic "Design a real-time notification system"`
|
||||
|
||||
**Round 1**:
|
||||
1. **Gemini**: Proposes WebSocket-based architecture with RabbitMQ message queue
|
||||
2. **Codex**: Critiques as overly complex for MVP. Suggests Server-Sent Events (SSE) for simplicity (one-way notifications). Questions RabbitMQ necessity, proposes simpler Redis Pub/Sub
|
||||
3. **Claude**: Synthesizes views:
|
||||
- **Plan A (Gemini)**: WebSockets + RabbitMQ (highly scalable, complex)
|
||||
- **Plan B (Codex)**: SSE + Redis (simpler, less overhead)
|
||||
- **Open Question**: Is bi-directional communication critical, or is simplicity priority?
|
||||
4. **User Action**: Opts for another round to explore trade-offs
|
||||
|
||||
**Round 2**:
|
||||
1. **Gemini**: Analyzes synthesized document. Notes that if features like "user is typing" indicators are roadmapped, WebSockets better long-term. Drafts plan starting with SSE/Redis but designing for easy migration
|
||||
2. **Codex**: Reviews migration plan. Reasons that migration itself could be complex. If feature set likely to expand, starting with WebSockets using managed service might be best cost/benefit
|
||||
3. **Claude**: Synthesizes new discussion:
|
||||
- **Consensus**: Simple SSE/Redis too short-sighted
|
||||
- **Refined Options**:
|
||||
1. Phased approach (SSE → WebSocket) with clear migration plan
|
||||
2. Direct WebSocket with managed service (Pusher, Ably) to reduce ops overhead
|
||||
- **Recommendation**: Option 2 most robust and future-proof
|
||||
4. **User Action**: Agrees with recommendation, concludes discussion
|
||||
|
||||
**Final Output**: Planning document saved with:
|
||||
- Chosen architecture (Managed WebSocket service)
|
||||
- Multi-round reasoning
|
||||
- High-level implementation steps
|
||||
|
||||
### Example 2: Feature Design Discussion
|
||||
|
||||
**Command**: `/cli:discuss-plan --topic "Design user permission system" --rounds 2`
|
||||
|
||||
**Round 1**:
|
||||
1. **Gemini**: Proposes RBAC (Role-Based Access Control) with predefined roles
|
||||
2. **Codex**: Suggests ABAC (Attribute-Based Access Control) for more flexibility
|
||||
3. **Claude**: Synthesizes trade-offs between simplicity (RBAC) vs flexibility (ABAC)
|
||||
|
||||
**Round 2**:
|
||||
1. **Gemini**: Analyzes hybrid approach - RBAC for core permissions, attributes for fine-grained control
|
||||
2. **Codex**: Reviews hybrid model, identifies implementation challenges
|
||||
3. **Claude**: Final plan with phased rollout strategy
|
||||
|
||||
**Automatic Conclusion**: Command concludes after 2 rounds as specified
|
||||
|
||||
### Example 3: Problem-Solving Discussion
|
||||
|
||||
**Command**: `/cli:discuss-plan --topic "Debug memory leak in data pipeline" --task-id ISSUE-042`
|
||||
|
||||
**Round 1**:
|
||||
1. **Gemini**: Identifies potential leak sources (unclosed handles, growing cache, event listeners)
|
||||
2. **Codex**: Adds profiling tool recommendations, suggests memory monitoring
|
||||
3. **Claude**: Structures debugging plan with phased approach
|
||||
|
||||
**User Decision**: Single round sufficient, concludes with debugging strategy
|
||||
|
||||
## Consensus Mechanisms
|
||||
|
||||
**When to Continue:**
|
||||
- Significant disagreement between models
|
||||
- Open questions requiring deeper analysis
|
||||
- Trade-offs need more exploration
|
||||
- User wants additional perspectives
|
||||
|
||||
**When to Conclude:**
|
||||
- Models converge on solution
|
||||
- All key questions addressed
|
||||
- User satisfied with plan depth
|
||||
- Maximum rounds reached (if specified)
|
||||
|
||||
## Comparison with Other Commands
|
||||
|
||||
| Command | Models | Rounds | Discussion | Implementation | Use Case |
|
||||
|---------|--------|--------|------------|----------------|----------|
|
||||
| `/cli:mode:plan` | Gemini | 1 | ❌ NO | ❌ NO | Single-model planning |
|
||||
| `/cli:analyze` | Gemini/Qwen | 1 | ❌ NO | ❌ NO | Code analysis |
|
||||
| `/cli:execute` | Any | 1 | ❌ NO | ✅ YES | Direct implementation |
|
||||
| `/cli:codex-execute` | Codex | 1 | ❌ NO | ✅ YES | Multi-stage implementation |
|
||||
| `/cli:discuss-plan` | **Gemini+Codex+Claude** | **Multiple** | ✅ **YES** | ❌ **NO** | **Multi-perspective planning** |
|
||||
|
||||
## Best Practices
|
||||
|
||||
1. **Use for Complex Decisions**: Ideal for architectural decisions, design trade-offs, problem-solving
|
||||
2. **Start with Broad Topic**: Let first round establish scope, subsequent rounds refine
|
||||
3. **Review Each Synthesis**: Claude's synthesis is key decision point - review carefully
|
||||
4. **Know When to Stop**: Don't over-iterate - 2-3 rounds usually sufficient
|
||||
5. **Task Association**: Use `--task-id` for traceability in workflow
|
||||
6. **Save Intermediate Results**: Each round's synthesis saved automatically
|
||||
7. **Let Models Disagree**: Divergent views often reveal important trade-offs
|
||||
8. **Focus Questions**: Use Claude's open questions to guide next round
|
||||
|
||||
## Breaking Discussion Loops
|
||||
|
||||
**Detecting Loops:**
|
||||
- Models repeating same arguments
|
||||
- No new insights emerging
|
||||
- Trade-offs well understood
|
||||
|
||||
**Breaking Strategies:**
|
||||
1. **User Decision**: Make executive decision when enough info gathered
|
||||
2. **Timeboxing**: Set max rounds upfront with `--rounds`
|
||||
3. **Criteria-Based**: Define decision criteria before starting
|
||||
4. **Hybrid Approach**: Accept multiple valid solutions in final plan
|
||||
|
||||
## Notes
|
||||
|
||||
- **Pure Discussion**: This command NEVER modifies code - only produces planning documents
|
||||
- **Codex Role**: Codex participates as reasoning/critique tool, not executor
|
||||
- **Resume Context**: Codex maintains discussion context via `resume --last`
|
||||
- **Priority System**: Ensures Gemini leads analysis, Codex provides critique, Claude synthesizes
|
||||
- **Output Quality**: Multi-perspective discussion produces more robust plans than single-model analysis
|
||||
- Command patterns and session management: see intelligent-tools-strategy.md (loaded in memory)
|
||||
- Output routing details: see workflow-architecture.md
|
||||
- For implementation after discussion, use `/cli:execute` or `/cli:codex-execute` separately
|
||||
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user