01
Core test
Can the client read the complete file, collect inputs first, follow Steps/Never and return Done means?
Compatibility is not one yes-or-no claim. This matrix separates the core file, attachments, tools and rendering so you know what a client can actually do.
Use “Tested” only for evidence recorded by AILoadout.
| AI client | Core file | Attach .md | Tools / MCP | Visual output | Evidence |
|---|---|---|---|---|---|
| ChatGPT | Tested | Conditional | Account-dependent | Native when available; Markdown fallback | Core · 23 Jul 2026 |
| Claude | Tested | Supported | Connected support | Custom visual / Artifact / fallback | Core · 23 Jul 2026 |
| Gemini | Tested | Supported | Surface-dependent | Charts for some tasks; fallback | Core · 23 Jul 2026 |
| Claude Code | File-compatible | Workspace file | Installable MCP | Markdown / Mermaid / HTML / Artifact | Connected path |
| Claude Cowork | File-compatible | Workspace file | Connector-dependent | Custom visual + fallback | Connected path |
| Codex | Expected | Workspace file | Tool-dependent | Markdown / artifacts by surface | Not yet in test corpus |
| Cursor | Expected | Workspace file | MCP-dependent | Markdown / code artifacts | Not yet in test corpus |
| Windsurf | Expected | Workspace file | MCP-dependent | Markdown / code artifacts | Not yet in test corpus |
A model name alone is not a compatibility result.
Can the client read the complete file, collect inputs first, follow Steps/Never and return Done means?
Can the declared tool be configured, scoped, approved and verified without inventing a capability?
Can it produce the preferred format or preserve every required component in structured Markdown?