Speak French to someone who only understands Japanese and you will not get far. The same is true of AI models, and the two labs that make the leading ones just proved it in public. Anthropic's Applied AI team put out a free 40-technique workshop built heavily around structured, XML-tagged prompting for Claude. OpenAI's official Codex Prompting Guide teaches a completely different structure: goal, context, constraints, done-when, no XML in sight. These are not competing opinions. They are two labs each telling you, directly, that their model has its own dialect.
Different models, different languages
Every major AI lab trains its models differently, which produces real differences in what each one responds to best:
- Claude (Anthropic). Responds well to XML-style structure, tags like
<document>and<instructions>, and a clearly defined persona. - Codex / ChatGPT (OpenAI). Built around the goal, context, constraints, done-when structure from OpenAI's own guide, with plain markdown over XML.
- Gemini (Google). Handles very large context windows well, and benefits from explicit grounding instructions.
- Cursor. Not just an LLM call. An IDE agent that needs file paths, codebase conventions, and awareness that it is operating inside a live development environment.
The danger of write once, run anywhere
Using one generic prompt across every tool means settling for the lowest common denominator. You are not using Claude's structured-tag parsing, or Codex's goal/context/constraints framing, or Cursor's codebase awareness. That is performance left on the table, and it is the exact gap the labs' own guides are trying to close, one tool at a time, for people willing to read two very different manuals.
Tool-specific compilation
This is a core part of what Kosmo does. You write your intent once, in plain language, and pick your target tool. Kosmo's compiler formats the result in that tool's native dialect: structured tags for Claude, the goal/context/constraints/done-when shape for Codex, environment-aware context for Cursor. You do not need to have read either lab's guide. Kosmo already has. Your prompt always knows its audience.
Frequently asked questions
Can I really not just copy the same prompt between Claude and ChatGPT? You can, but you will underperform both. The labs themselves publish different structural guidance for a reason: the models were trained to respond best to different shapes.
How does Kosmo know which dialect to use?You pick the target tool when you compile, and Kosmo formats the output against that tool's current, documented conventions, not a guess based on old training data.