2 min lesson
The authentic-usage test
Rebuild the parts of "The authentic-usage test", then say why each one matters.
Step 1 of 2
The loop detects superficial familiarity within minutes. Generic praise - “I love how fast it is” - signals elimination, because a routing-and-inference candidate who only skimmed the product can't reason about the path they'd be hired to own.
Specificity is the whole test. Name the surfaces, describe a real task on a real repo and say what worked and what frustrated you. For this role especially, your critiques should land on the inference experience: a Tab completion that arrived late, an Agent run that stalled mid-stream, a model swap that changed answer quality.
- Tab
- Inline multi-line completion on a sub-100ms budget - the most latency-sensitive surface
- Cmd+K
- Inline edit/generate on a selection - interactive, expects a fast first token
- Agent / Composer
- Multi-step, multi-file edits - long-running, tolerant of latency but sensitive to mid-stream failure
- Chat
- Conversational with codebase context - streamed, quality-sensitive across model choice
Each surface has a different latency/cost/quality profile - the gateway you'd own routes for all of them.
“I run Cursor daily on a TypeScript monorepo. Tab is the surface I'd defend hardest - when first-token latency creeps over ~120ms it stops feeling like autocomplete and I start typing past it. Agent is where I lean on it for refactors and the failure I notice most is a run dying mid-stream with partial edits, which is a failover and idempotency problem more than a model problem.”