Skip to main content
Back
45 XP

Lesson 8 · 10 min

Take-homes, behavioral, and a self-assessment rubric

What take-homes look like in 2026, the behavioral patterns that come up in AI eng loops, and a rubric you can use to score yourself before the real loop.

Take-homes

The 2026 take-home patterns:

  • "Build a small RAG pipeline on this corpus and answer these 10 questions." Time-boxed at 4-6 hours. Submitted as a repo + a README. Graded on: works at all (table stakes); chunking + retrieval rationale; eval discipline; observability; clear README.
  • "Write a prompt that solves these 20 cases. We'll grade on a held-out 20." Time-boxed at 2-3 hours. Graded on accuracy, robustness, and the prompt's clarity — not its length.
  • "Here's a broken prompt + eval set. Fix it." Time-boxed at 2 hours. Graded on triage discipline (per-case fails read, bucketed) and the per-case diff between your fix and the baseline.
  • "Build a small agent for this task." Time-boxed at 6-8 hours. Graded on tool design, bounded loops, and trajectory eval.

All four reward eval discipline more than raw cleverness. Submit a repo with a 50-line eval runner and per-case scoring.