I want to tell you about the specific flavor of wrong that is being marked incorrect on a certification quiz using your own reasoning.
Not wrong because you misunderstood. Not wrong because you guessed. Wrong because the question was broken, you applied a clean mental model correctly, and the answer key had a different interpretation of the same ambiguous word.
That flavor of wrong has a name. It’s called Tuesday.
The Setup
I’ve been working through Anthropic Academy’s MCP courses — Intro and Advanced Topics back to back. MCP, for the uninitiated, is Model Context Protocol: the standard that lets Claude connect to external tools and data sources. Think of it as USB-C for AI. Universal connector. Any AI, any tool, one standard.
It has three primitives:
Tools — actions. Things that do something. Call an API. Write to a database. Side effects welcome.
Resources — data. Things that are something. Files, records, documents Claude can read.
Prompts — reusable templates. Pre-built workflows surfaced to users as options they can invoke.
I understood this distinction from lesson one. It stayed stable through every example, every edge case, every follow-up question I threw at it.
Then came the quiz.
The Crime
Question: a user clicks a button labeled “Summarize Document.” Which MCP primitive handles that?
I read it. Thought about it. Applied my mental model.
A user clicking a button to trigger a workflow — that’s a user-initiated action. That’s a Tool. Tools do things. The document itself would be a Resource. But the click, the trigger, the “go summarize this now” — Tool.
I ruled out Prompts specifically because of my reasoning: users control when to start the workflow. That’s the characteristic of a Tool. Prompts are templates. Pre-built patterns. Not the same as a trigger.
Selected Tool. Confidently.
Quiz: ❌ Prompts.
I read the explanation.
The quiz’s own justification for why Prompts was correct:
“Users control when to start the workflow.”
I want you to read that again.
The quiz marked me wrong. Then explained why I was wrong. Using the exact phrase I used to rule out the answer it wanted.
I have been gaslit by a multiple-choice question. This is a new experience for me, and I do not recommend it.
What Actually Happened
After the initial indignation passed — which took a while — here’s what I think the quiz was actually going for.
A “Summarize Document” button in a UI is a surfaced Prompt template. The user sees a button, but under the hood it’s a pre-built Prompt that gets invoked when clicked. The quiz wanted me to think about which primitive exposes that workflow to the user, not which primitive executes the action.
Legitimate distinction. Genuinely useful thing to understand.
Also not what the question said.
“Handles that” is doing too much work in that sentence. Handle could mean triggers. Could mean executes. Could mean surfaces. If the question had said “which MCP primitive surfaces this workflow in the UI” — Prompts, no debate. If it had said “which primitive executes the action when clicked” — Tool, also no debate.
It said neither. And so here we are.
I built the right mental model. Applied it consistently. Got marked wrong by a question that wasn’t precise enough to have one correct answer.
The One Thing The Course Forgot to Mention
Practical note for anyone taking this course:
The material tells you to run mcp dev mcp_server.py without explaining where the mcp command comes from.
It comes from pip install mcp[cli].
One line. Not in the materials. You find out when the command doesn’t run, and you go looking. Consider this your advance warning. You’re welcome.
What Actually Clicked
The Tool/Resource/Prompt distinction — quiz drama aside — is genuinely clean once it settles.
Tools are verbs. Resources are nouns. Prompts are templates.
If you’re ever building with MCP and asking “which primitive is this” — ask whether it does something, is something, or is a reusable pattern for something. That’s the whole model.