Tailored from the JD. Private GitHub repo. Webhook-graded. Ships only to candidates whose interview clears the threshold.
Trusted by recruiting teams shipping audit-grade hiring
01 · Topic
Claude Sonnet generates a per-candidate challenge — not a static problem set. The topic respects the candidate's stated stack and the role's must-haves. Repo scaffold lands as README.md (challenge spec), CLAUDE.md (AI session log preservation rules), and .github/workflows/notify.yml (HMAC-signed webhook back to the engine).
02 · Submission
Candidate works in a private repo; pushes trigger a workflow that POSTs to /api/challenge/webhook with a per-repo HMAC-SHA256 signature. Status flows pending_github → active → submitted → checking → evaluated. Deadline expiry triggers automatic evaluation; the cron (check-challenge-deadlines) sweeps every 10 minutes for orphans.
03 · Evaluation
Functionality 30% · Code quality 20% · Engineering 20% · AI collaboration 20% · Time efficiency 10%. The Challenge Eval Agent (Agent SDK) reads repo contents via the GitHub API + commit history + .claude/ AI logs + the challenge spec + the JD. Score → recommendation: ≥70 hire / 50–69 maybe / <50 pass. Report email lands in the hiring manager's inbox.
Gate
Topic generator + autonomous evaluator dispatch a PR-grade challenge only to candidates that clear the rubric threshold. The grader's own provenance is logged the same way the interview is.