適切なプリミティブを選ぶ

You know tools, resources, and prompts. Now the skill is choosing the right one for each feature you want to expose. 🧭

The clearest divider is control. Tools are model-controlled, resources are application-controlled, and prompts are user-controlled.

If a feature does something, like sending mail or writing a row, it is almost always a tool. Action is the tool's signature trait.

If the model only needs to read something, a file, record, or setting, reach for a resource. No side effects, just context.

If you want a one click, repeatable way to start a task, that is a prompt. The user triggers it and fills in a few inputs.

Fetching live weather usually fits a tool, since the model decides when it needs fresh data and passes a city argument.

Your project's README is static context the AI should see. That is a textbook resource, attached by the app, not run by the model.

A reusable summarize this command the user picks from a menu is a prompt. It templates the instruction so every summary starts the same way.

Real servers combine all three: a resource feeds context, a prompt frames the request, and a tool carries out the resulting action.

Unsure between tool and resource? Ask whether it changes anything. If yes, it is a tool; if it only informs, it is a resource.

Before coding, list each feature and tag it tool, resource, or prompt. That quick mapping keeps your server clean and predictable.

Pick the right primitive for this case.

Well done! Ask who controls it and whether it acts: tool for actions, resource for read-only data, prompt for saved requests. ✅