Prompt engineer
is not the job.
Being able to prove it improved is.
Anyone can make a demo answer one question well. Our GenAI engineers ship retrieval, an evaluation set, guardrails and tracing — so you can tell whether this week's change helped or quietly broke something.
One row got worse. That is the point.
A prompt change that improves six categories and degrades a seventh is a normal Tuesday. Without a scorecard you ship it and find out from a customer in March.
“Can you show me your eval set?” is the single best interview question in this discipline. Ask it before you ask about model architectures.
“Stays inside scope” fell four points and latency regressed five. Both were traced to the same longer system prompt, and both were fixed before release — because the scorecard runs in CI, not at the end of the quarter.
Six components, only one of which is a prompt
Retrieval, properly
Chunking that respects document structure, hybrid search, a re-ranker, and metadata filters applied at query time rather than after.
Evaluation as a build artefact
A graded set written with your subject-matter experts, run in CI. A regression blocks the release like any failing test.
Guardrails with teeth
Scope checks, PII redaction, output schemas, and a refusal path that is tested rather than hoped for.
Tracing and cost control
Every prompt, retrieval and tool call logged with token counts and cost, so “why did last month cost $9k” has an answer.
Model portability
Providers behind one interface. When a better or cheaper model lands, switching is a config change and a re-run of the eval set.
Agent design where it earns it
Tool definitions, permissioned write actions, and an authorisation model designed before the agent — not bolted on after a demo.
The most inflated titles on the market
This discipline is three years old, so “eight years of experience” is not available and anyone claiming it is rounding. We band on adjacent depth — search, ML or distributed systems — plus demonstrable GenAI delivery.
Mid-level
3–5 yrs$45–55/hrBuilds RAG pipelines and prompts against an existing architecture. Comfortable with the SDKs; still learning evaluation.
Senior
5–8 yrs$55–70/hrOwns the retrieval design, the eval set and the guardrails. Can explain why an answer was wrong, from the trace.
Lead
8+ yrs$70–95/hrSets the platform, the model strategy and the authorisation model for agents. Also the person who says a use-case is not viable.
Tell us the question users will actually ask
Twenty real examples is enough for us to say whether retrieval will answer them, what it costs, and whether you need one engineer or a team.