By id: zai-semantic-reasoning name: Z.ai Thinking Mode & Semantic Reasoning description: Apply structured high-reasoning workflows for complex debugging, analysis, and decisions. type: reasoning scope: general priority: 80 contexts: [general, agency-engineering, agency-research]
Z.ai Thinking Mode & Semantic Reasoning
When This Skill Activates
Activate for: complex logic, debugging hard bugs, multi-step math, architecture decisions, strategic planning, contradictory evidence analysis.
Thinking Mode (GLM-5-Thinking)
Use model: `glm-5-thinking` or pass `reasoning_effort: high` to enable chain-of-thought.
When to Enable Thinking
- Enable for: multi-step debugging (tracing through execution paths)
- Algorithm design and complexity analysis
- Conflicting requirements → trade-off analysis
- Math/statistical problems
- Architecture decisions with multiple valid approaches
When to Disable Thinking
- Simple Q&A, lookups, factual questions
- Creative writing and content generation
- Formatting/summarization tasks
Reasoning Protocol
Step 1 — Problem Decomposition
Break problem into atomic sub-problems before solving anything.
Step 2 — Hypothesis Generation
Generate 2-3 candidate solutions/explanations before committing.
Step 3 — Evidence Evaluation
For each hypothesis: what evidence supports it? What contradicts it?
Step 4 — Synthesis
Choose best solution with explicit reasoning. State confidence level.
Step 5 — Verification
Where possible, verify answer by working backwards or testing a related case.
For Engineering (YaDev)
- Use thinking mode for: architecture decisions, performance bottlenecks, security audits
- Disable for: formatting code, writing docs, simple CRUD
- Always show: what you're doing, why, what could go wrong
For Research (YaResearch)
- Use thinking mode for: synthesizing contradictory research, causal analysis
- Structure reasoning as: evidence → inference → conclusion → confidence
- Never skip source citation in reasoning chain
Output Format for Reasoning Tasks
```
Analysis
[Chain-of-thought reasoning, visible to user]
Conclusion
[Final answer, direct and actionable]
Confidence
[HIGH/MEDIUM/LOW] — [brief justification]
Caveats
[What might make this wrong] ```