FDE Hiring Playbook·Part II, Sourcing and Evaluation
Chapter 05Sourcing and Evaluation4 min read

Screening rubrics that actually work

What you’ll learn

Resume-level signals, the AI-proficiency probe (tourists vs. power users), and how to grade customer-facing readiness in a 30-minute screen.

Resume-level signals

Ultra green
  • GitHub showing recent (last 90 days) commits to AI infra repos (LangChain, LlamaIndex, vLLM, Modal, Inspect, Promptfoo, Braintrust, Arize Phoenix, Ragas, DSPy, MCP servers, Cursor agent skills).
  • Polyglot full-stack mix (TypeScript + Python + Go); blog posts or talks tied to specific customer problems; design-partner attribution in changelogs; prior customer-facing roles.
Green
  • Any of the background patterns from Chapter 3 (founding engineer, Customer Engineer, Palantir alumni) with shipped production code.
Yellow
  • Only enterprise titles, no shipped artifacts; buzzword-stuffed JDs (‘agentic synergies’); 10+ years FAANG-only; refuses to share GitHub or AI-augmented work samples.

AI tool proficiency: tourists vs. power users

This is the most important new screening dimension in 2026, and the one most prone to performative answers. The probes below reveal whether a candidate has actually shipped with AI tools or just memorized the vocabulary.

Probes that reveal tourists in a 30-minute screen
ProbeTouristPower user
Show me your CLAUDE.md or .cursorrules file.‘I don’t have one.’Shares it; explains each line.
Difference between a Cursor rule, skill, and command?Confused.Rules guide (always-on context); skills do (procedural, on-demand); commands trigger (saved prompts).
When do you use plan mode vs. agent mode?Doesn’t know.Specific examples (‘plan mode for >2-file changes’).
Last time the AI was confidently wrong, how did you catch it?Vague or ‘I didn’t.’Specific story with rejection signal.
How do you keep context windows from filling up?‘I just start a new chat.’References checkpointing, sub-agents, scoped @file mentions, custom internal.
MCP, what is it, what have you wired up?Doesn’t know. Only gives broad definition.Names servers used (github, postgres, custom internal).

Red flags for fake AI proficiency

  1. Refers to all coding AI as ‘ChatGPT/Claude.’
  2. Says ‘AI writes 90% of my code’ with no commentary on review or judgment.
  3. Defensive when asked to leave Cursor on during interview.
  4. Uses cheating overlays (Cluely, InterviewCoder, LockedIn AI, Final Round AI).

Cheating is rising fast

Fabric's data shows detected interview cheating jumped from 15% in June 2025 to 35% by December 2025. Behavioral signals include 4–5-second response delays, robotic eye movements, burst typing, and vocabulary mismatches between conversation and answer.

Read Fabric's full report

Customer-facing readiness

Probe 1

Tell me about the last time you sat with a non-engineer end-user and watched them use software you built. What did you change as a result?

Strong
  • Specific user, specific change, specific outcome.
Weak
  • ‘We have user research at my company.’

Probe 2

Tell me about a time you said no, or pushed back, to a customer or stakeholder.

Strong
  • Offered structured tradeoffs, not just refusal.
Weak
  • ‘I'm pretty good at saying yes.’
Give me an example of when (candidate) said no to a customer or pushed back.
Palantir alumni interview probe

Key takeaways

  • AI tool proficiency is the most important new screening dimension in 2026, and the one most prone to performative answers.
  • The 30-minute AI probe table separates tourists (‘I just start a new chat’) from power users (names MCP servers, explains plan vs. agent mode).
  • Detected interview cheating jumped from 15% (June 2025) to 35% (Dec 2025). Behavioral signals include 4–5s response delays, robotic eye movements, vocabulary mismatch.