How to hire a prompt engineer in 2024 (without getting burned)
A practical rubric for evaluating prompt engineering work, plus red flags to avoid.
The phrase 'prompt engineer' covers a huge range. On one end you have someone who can craft a clever one-shot that gets ChatGPT to write better blog posts. On the other you have engineers building structured output pipelines, eval harnesses and multi-step tool-calling agents in production. Both are valid; only one of them solves the business problem you usually have.
When you hire, write the brief like an engineering brief. State the inputs, the desired outputs, the failure modes you've seen, and the eval criteria. Ask the candidate to show you a small repo or notebook where they iterated on a prompt with measurable improvements between versions. If they only ever paste prompts into a chat window, that's a red flag.
Don't pay for a 'master prompt'. Pay for a system: prompt + structured output schema + eval set + a short loom showing how to update each piece when the model behind it changes. That's the deliverable that survives the next GPT release.