Agent Skills
› modu-ai/moai-adk
› moai-workflow-thinking
moai-workflow-thinking
GitHub通过 MCP 工具实现分步推理，用于复杂问题分析、架构决策及技术选型。
Trigger Scenarios
显式使用 --deepthink 标志
需要分解复杂问题或进行多步骤规划
Install
npx skills add modu-ai/moai-adk --skill moai-workflow-thinking -g -y
SKILL.md
Frontmatter
{
"name": "moai-workflow-thinking",
"effort": "high",
"license": "Apache-2.0",
"metadata": {
"status": "active",
"version": "2.0.0",
"category": "workflow",
"modularized": "false"
},
"triggers": {
"agents": [
"manager-strategy",
"manager-spec"
],
"phases": [
"plan"
],
"keywords": [
"sequential thinking",
"deepthink",
"deep analysis",
"complex problem",
"architecture decision",
"technology selection",
"trade-off",
"breaking change"
]
},
"description": "Sequential Thinking MCP for structured step-by-step analysis via --deepthink flag. Separate from UltraThink which is Claude's native extended reasoning mode. Use for multi-step analysis or architecture decisions.\n",
"allowed-tools": "Read, Grep, Glob, mcp__sequential-thinking__sequentialthinking",
"compatibility": "Designed for Claude Code",
"user-invocable": false,
"progressive_disclosure": {
"enabled": true,
"level1_tokens": 100,
"level2_tokens": 3000
}
}
Sequential Thinking MCP (--deepthink)
Structured step-by-step reasoning via mcp__sequential-thinking__sequentialthinking MCP tool.
CRITICAL: Three Distinct Reasoning Modes
MoAI has THREE independent deep analysis modes. They are NOT the same thing:
| Mode | Trigger | Mechanism | MCP Tool? | GLM Compatible? | Model |
|---|---|---|---|---|---|
--deepthink |
Explicit --deepthink flag |
Sequential Thinking MCP tool | YES — mcp__sequential-thinking__sequentialthinking |
NO — generates server_tool_use content type | Any |
ultrathink |
Keyword or auto-detection | Claude native extended reasoning (high effort) | NO — native to Claude | YES — no special content type | Any |
| Adaptive Thinking | Automatic on Opus 4.7 | Opus 4.7's only supported thinking mode | NO — built-in | YES | Opus 4.7 only |
Rules:
--deepthink→ ALWAYS invoke Sequential Thinking MCP. NEVER use for native reasoning.ultrathink→ ALWAYS use Claude's native extended reasoning. NEVER invoke Sequential Thinking MCP.- They can coexist:
ultrathink --deepthinkactivates BOTH modes independently. - Adaptive Thinking → Opus 4.7's built-in reasoning. Let the model adapt depth automatically based on task complexity.
- On Opus 4.7: do not hardcode a fixed reasoning budget. Adaptive Thinking overrides fixed-budget instructions.
Activation Triggers (--deepthink only)
Use Sequential Thinking MCP when --deepthink flag is explicitly present:
- Breaking down complex problems into steps
- Planning and design with room for revision
- Architecture decisions affect 3+ files
- Technology selection between multiple options
- Performance vs maintainability trade-offs
- Breaking changes under consideration
- Multiple approaches exist to solve the same problem
- Repetitive errors occur
Tool Parameters
Required Parameters:
thought(string): Current thinking step contentnextThoughtNeeded(boolean): Whether another step is neededthoughtNumber(integer): Current thought number (starts from 1)totalThoughts(integer): Estimated total thoughts needed
Optional Parameters:
isRevision(boolean): Whether this revises previous thinkingrevisesThought(integer): Which thought is being reconsideredbranchFromThought(integer): Branching point for alternativesbranchId(string): Branch identifierneedsMoreThoughts(boolean): If more thoughts needed beyond estimate
Usage Pattern
Step 1 - Initial Analysis:
thought: "Analyzing the problem: [describe problem]"
nextThoughtNeeded: true
thoughtNumber: 1
totalThoughts: 5
Step 2 - Decomposition:
thought: "Breaking down: [sub-problems]"
nextThoughtNeeded: true
thoughtNumber: 2
totalThoughts: 5
Step 3 - Revision (if needed):
thought: "Revising thought 2: [correction]"
isRevision: true
revisesThought: 2
thoughtNumber: 3
totalThoughts: 5
nextThoughtNeeded: true
Final Step - Conclusion:
thought: "Conclusion: [final answer]"
thoughtNumber: 5
totalThoughts: 5
nextThoughtNeeded: false
Guidelines
- Start with reasonable totalThoughts estimate
- Use isRevision when correcting previous thoughts
- Maintain thoughtNumber sequence
- Set nextThoughtNeeded to false only when complete
- Use branching for exploring alternatives
Common Rationalizations
| Rationalization | Reality |
|---|---|
| "I can think through this without sequential thinking, it is simple" | Simple problems often hide second-order effects. Structured thinking forces you to enumerate them. |
| "The thinking steps are just internal, I do not need to record them" | Unrecorded reasoning cannot be reviewed. The chain of thoughts is evidence for the conclusion. |
| "I already know the answer, the thinking framework is overhead" | Confirmation bias skips disconfirming evidence. The framework forces you to consider alternatives. |
| "Branching is overkill for this decision" | Decisions with more than one viable path benefit from explicit branch comparison, even briefly. |
| "I will use UltraThink instead, it is the same thing" | UltraThink is native extended reasoning. Sequential Thinking is MCP-based structured analysis. They serve different purposes and have different API compatibility. |
Red Flags
- Architecture decision made without documented reasoning chain
- Sequential thinking session ended with nextThoughtNeeded still true
- Thought chain has no revision steps despite encountering contradictions
- Branching not used when two or more viable alternatives were identified
- --deepthink flag confused with UltraThink (wrong tool for the job)
Verification
- Thinking chain has a clear totalThoughts estimate that was met or revised
- Final thought sets nextThoughtNeeded to false with a conclusion
- At least one revision step exists if contradictions were encountered
- Branching used when multiple alternatives were identified (show branch IDs)
- Conclusion references specific thought numbers as supporting evidence
Version History
- 4100d87 Current 2026-08-20 09:10
