Customizing an AI agent in 2026 is no longer one decision. It is nine. Claude Skills, Prompts, MCP servers, Claude Projects, CLAUDE.md, AGENTS.md, Cursor Rules, OpenAI Custom GPTs, and OpenAI AgentKit all do overlapping things in different places. The matrix below maps each one to where it lives, when it loads, and when to use it.
| Tool | Lives at | Loads when | Use it when |
|---|---|---|---|
| Claude Skill | Folder with SKILL.md, scripts, references | Claude matches your request to the description | You repeat a procedure 3+ times and want it portable |
| Prompt | A text snippet you paste | You paste it | One-off or rarely repeated work |
| MCP server | A running server you connect to | The connection is open and you use the tool | You need live data or to write to an external system |
| Claude Project | A workspace with files and instructions | You're inside the Project | The work is shared across a defined scope |
| CLAUDE.md | Root of a code repo | Claude Code is invoked in that repo | You want repo-specific conventions, paths, commands |
| AGENTS.md | Root of a code repo | Any agent that supports the standard reads it | You want cross-agent conventions |
| Cursor Rules | .cursor/rules folder | You're in Cursor | You want Cursor-specific behavior, not portable |
| OpenAI Custom GPT | OpenAI dashboard | You start a chat with that GPT | You want a packaged ChatGPT persona |
| OpenAI AgentKit | API-level agent definition | You call the agent endpoint | You're building production multi-agent systems on OpenAI |
Source: Product with Attitude. Tested April 2026.