Prompt Engineering Interview Rubric (2026)

Prompt engineering is no longer a side skill—it sits on the critical path for RAG, agents, and customer-facing copilots. A shared rubric keeps interviewers aligned when multiple teams assess the same candidate.

This playbook targets roles where prompts, tool schemas, and eval loops are weekly work—not pure research scientist tracks. Adapt weights if the hire is mostly MLOps or backend.

Do they clarify user intent, constraints, and success criteria before writing prompts?

Do they anticipate ambiguity and propose a default, clarifying question, or fallback policy?

Score higher when they separate model behavior from product policy (what the UI promises vs what the model can reliably do).

Do they structure prompts with clear roles, examples, and output formats instead of one long paragraph?

Can they describe a regression strategy—golden sets, diff review, A/B on latency/cost—when requirements change?

Look for trade-off language: when to add retrieval, when to fine-tune, when to add a deterministic guardrail.

Do they mention prompt injection, tool misuse, PII leakage, or jailbreak patterns relevant to your surface?

Do they balance verbosity with latency and cost for the intended channel (batch vs interactive vs voice)?

Strong candidates propose logging, human review queues, or rate limits—not only “better prompts.”

1 — Cannot frame the task; prompts are unstructured; no mention of evaluation or failure modes.

2 — Writes workable prompts but weak on iteration, metrics, or security; needs heavy coaching.

3 — Solid product sense; proposes eval harness and guardrails; explains trade-offs clearly.

4 — Teaches the room: systematic eval design, cost/latency awareness, and crisp handoff to engineering/MLOps.

Give a realistic brief (support bot, document Q&A, internal search) with 20–30 minutes to draft prompts plus a mini eval plan.

Optional twist: change a constraint mid-exercise (new locale, stricter latency, banned tools) and observe adaptation.

Debrief on what they would ship in week one vs defer—signals product maturity.

Combine a live exercise (60%), a short take-home or portfolio walkthrough (30%), and culture/communication (10%). Weight safety higher for customer-facing roles.

Ask for one war story: a prompt that regressed in production and how they detected and fixed it.

For mid-level IC roles, require 3+ on problem framing and iteration, and no 1 on safety. Seniors should trend toward 4 on at least two dimensions.

If you are hiring your first LLM engineer, a strong 3 with backend or data skills often beats a theoretical 4 with no shipping history.