Why We Need a Multi-Model Adapter

Different models are trained on different corpora, aligned with different methods, and use different reasoning paradigms. A prompt or skill that's optimal for Claude is rarely optimal for GPT, Gemini, DeepSeek-R1, or Kimi. Writing one prompt for all of them means only one model performs well — the others degrade in quality or output stability. ## Where the Differences Come From | Dimension | Description | | --- | --- | | Training data | Anthropic used lots of XML-styled data — Claude is highly sensitive to tag structure. OpenAI prefers Markdown instructions | | Alignment | RLHF / DPO / Constitutional AI — each favors different prompt style (strict vs minimal) | | Reasoning paradigm | Native reasoning models (o1 / R1 / QwQ) produce their own CoT — adding "think step by step" **hurts** quality | | Context length | Kimi / Gemini have huge context → inline entire docs. Claude → progressive disclosure works better | | Tool protocol | Anthropic tool_use / OpenAI function calling / Google function_declarations / Chinese platform DSL all differ | | Language preference | Chinese models handle direct Chinese best; international models prefer Markdown + English | ## The Cost of "One Prompt Fits All" Common failure modes when swapping models: • XML prompt written for Claude → dropped into GPT → quality drops, structure unstable • "Let's think step by step" written for GPT → dropped into o1 / R1 → reasoning models get misled, waste reasoning budget • Claude's progressive-disclosure skill → Kimi → wastes Kimi's huge context ability • Chinese prompt unchanged → Gemini → noticeably worse than English equivalent • Chinese workflow platforms (Coze / Bailian / Qianfan) need DAG DSL → a flat prompt cannot be deployed directly > **WARNING**: "Barely runs" ≠ "performs optimally". A production-grade prompt/skill system must render differently per model. ## PromptMan's Approach PromptMan ships a "Canonical Skill Spec + Model Adapter" system: 1. **Canonical (vendor-neutral) middle layer** — Describe identity, goal, steps, inputs, output, constraints, examples, tools — without any vendor-specific syntax 2. **Adapter layer** — Every mainstream model (Claude / GPT / Gemini / Qwen / DeepSeek / Kimi / GLM / ERNIE / Doubao + reasoning variants) has a dedicated adapter that renders to its best-practice format 3. **Unified entrypoint** — renderSkill(spec, family) and a REST API — render one spec to all models in a single call 4. **Playground** — Visit /adapters to compare 12 model renderings live in the browser