Once you've picked an entry point, the next decision is which features to turn on. The right mental model is: task need → feature need, not "turn everything on by default."
| If the task needs... | Reach for... |
|---|---|
| Current, real-time, or post-cutoff information | Web search |
| Deeper reasoning on a hard, ambiguous problem | Extended / adaptive thinking |
| A durable, reusable context for a recurring task | Projects |
| Output the user will edit, reuse, or hand off | Artifacts |
| Data manipulation, calculations, file generation | Code execution / file creation |
| Data that lives in another system (Drive, Slack, etc.) | Connectors |
The failure mode to watch for isn't under-using features — it's over-using them. Turning on web search for a task with no time-sensitivity, or generating an Artifact for a two-sentence answer, adds latency and noise without adding value. Feature selection is itself a judgment call the certification tests.