What they do

The candidate lands in a Node.js monorepo they've never seen. The brief: add per-tenant rate limiting to the REST API without breaking three downstream services. They spend the first seven minutes reading — not prompting. They grep for middleware patterns, trace the request lifecycle, find where auth tokens are verified. When they finally open a conversation with Claude Code, their first message is a 90-word specification with explicit constraints: sliding window, Redis-backed, reject at 429 not 503, propagate the Retry-After header. They don't ask Claude what to do. They tell it exactly what to build.

What Promptster captures

What Promptster captures that you'd miss: the seven minutes of deliberate, pre-prompt exploration. Most tools see only the AI conversation. Promptster captures the Read sequence (routes.ts → middleware/auth.ts → lib/redis.ts → tests/integration/api.test.ts) that preceded the first prompt. It records that the candidate read the test suite before touching a single file. It flags a decision event at 14:22 when they manually rewrote the Redis key schema that Claude proposed — because the AI's version would've broken multi-region reads. Prompt quality score: 9.1/10. Human attribution: 81%.

How this changes the hire

How this changes your decision: you were prepared to debate whether to advance them based on the final PR. Now you can see they caught a distributed-systems footgun that your senior reviewer might have missed on code review. You advance them to final-round without a phone screen, because the telemetry already answered your architectural competency questions. The next candidate submitted faster and had cleaner code. Their telemetry showed 23 prompt-fire-and-hope cycles, zero exploration, and 94% AI attribution. You pass.