China Model Selection Guide

China Model Selection Guide

Follow this flow to recommend models. Load references/china-model-selection-guide.md for the full Chinese playbook, scenarios, strengths, and prompt templates.

Quick Triage

Answer two questions first, then give a primary pick.

1. Identify core input type

• Visual-first input (UI mockups, screenshots, sketches): prefer Doubao-Seed-2.0-Code
• Very long text or many files (dozens of docs, full codebase): prefer Kimi-K2.5
• Structured engineering prompts (clear coding requirements, Shell commands): prefer GLM-5 or MiniMax-M2.5

2. Identify task complexity

• Complex reasoning or autonomous planning (system design, codebase refactor): prefer GLM-5
• Cross-language engineering (Python/C++, Java/Go): prefer MiniMax-M2.5
• Clear task but heavy execution (UI-to-code, template generation): prefer Doubao-Seed-2.0-Code

Tie-Break Rules

When multiple models fit, decide in this order.

1. Satisfy hard constraints first: vision, long-context, cross-language, agentic planning
2. Then compare cost and latency: pick better price/performance at similar quality
3. Finally split by phase: allow multi-model routing inside one project

Composite Task Routing

Use this default pipeline.

1. Planning: GLM-5 for architecture, decomposition, interfaces, schema decisions
2. Build:

• Frontend and visual replication: Doubao-Seed-2.0-Code
• Backend scripts, cross-language tasks, terminal automation: MiniMax-M2.5

3. Integration debugging: route hard cross-module issues back to GLM-5
4. Documentation handoff: send codebase and large document sets to Kimi-K2.5

Output Format

Always include these in recommendations.

1. Decision: primary model + fallback model
2. Rationale: map to input type, complexity, and constraints
3. Risks: likely weak points and rollback strategy
4. Execution: a ready-to-use prompt draft

References

• Full guide and examples (Chinese): references/china-model-selection-guide.md