Case study

How this lab was designed

A hiring manager should be able to take the lab, open this page, and decide whether I think like their education team. This is the brief I would have written internally.

Audience

Builders who already chat with Claude. They do not need a model explainer. They need a map of a product family that outgrew one window. The same person shows up at Cursor, OpenAI, and any company shipping an agent next to a chat box.

Job to be done

Leave able to choose Chat, Cowork, Code, or Chrome by output type. Secondary job: notice the shared layer (Skills, Connectors, memory) so they stop re-teaching the model in every surface.

Constraints

Twenty minutes, self-serve, no account.
No live Claude, no API keys, no login wall. Voice walks each section.
Skimmable in an eight-minute hiring pass.
Must still work as a 45-minute facilitated workshop.

Why this sequence

Overview first so the product family has names. Map assigns each surface a job on one Flappy Bird build. One module per surface so “when not” is as loud as “when.” Connect is the assessment: spec, implement, playtest, write the README. Shared is the transfer — the layer follows you to the next product you teach.

What we refused

Feature laundry lists. They test memory, not judgment.
A live API demo. It breaks, it needs keys, and it looks like an engineering sample.
A second track for Cursor or ChatGPT in v1. Dilutes the case study.
Video-first. Hiring managers do not watch a course. They click.

What we would measure

Time to first correct surface: On the chooser and in-module checks, how quickly someone names the right reach.
When-not accuracy: Wrong-surface answers are the interesting misses. Chat-for-everything is the default failure.
Workshop completion: In a live room: did pairs finish two surfaces plus the connect task in 45 minutes.
Ticket deflection: If this shipped internally: fewer “which Claude do I open?” questions in support and Slack.

What transfers

Cursor has Chat, Agent, terminal, and browser. OpenAI has ChatGPT, Codex, Atlas, and custom GPTs. The names change. The education problem does not: teach the map, not the model. Claude is the worked example because it is the clearest multi-surface product you can teach without an account.

45-minute workshop run of show · Take the lab