Documentation

Model-by-Model Differences

International Models

ModelPreferred FormatContextReasoningKey Notes
Claude (Anthropic)XML200K+No (has extended thinking)Wrap each semantic block in its own <tag>; also produce SKILL.md (progressive disclosure)
GPT non-reasoning (4o/5)Markdown128K+NoHigh few-shot payoff; use Structured Outputs for JSON
o-series / GPT-5-ThinkingPlain200K+YesStrip CoT prompts; currently no system role → merged into first user turn
Gemini 1.5/2.0/2.5Markdown1M-2MNoNative long-context — inline knowledge directly; function_declarations types must be UPPERCASE

Chinese Models

ModelPreferred FormatContextReasoningKey Notes
Qwen 2.5 / 3Markdown128KNoOpenAI-compatible; native Chinese; Bailian agent/workflow support
QwQ / Qwen3-ThinkingPlain128KYesAligns with o1/R1 style; no CoT prompts
DeepSeek V3Markdown128KNoOpenAI-compatible API; strong on code & math
DeepSeek R1Plain128KYesAuto-produces <think>; never add 'think step by step'
Kimi (Moonshot)Natural Chinese128K-2MNoHuge context; paste long docs straight into system
Zhipu GLM-4Markdown128KNoStrong Chinese writing; OpenAI-compatible
ERNIE (Baidu)Markdown128KNoBest for China-local context; Qianfan prefers workflow DSL
Doubao / Coze (ByteDance)Markdown + DSL128KNoFast & cheap; Coze = persona + plugins + KB + workflow

NOTE

PromptMan's adapters encode these characteristics — you don't need to memorize them manually.

Continue Reading