Prompt engineering
Hire prompt engineers, then find out the model was never the problem
A prompt engineer can improve wording. Most inconsistent AI features fail on evidence and measurement, not phrasing, and that needs a wider skill set than one title covers.
What you are getting
4 things that decide this
- 01We do not staff prompt engineering as its own role. It is one skill inside LLM engineering, alongside evals, structured output and retrieval.
- 02Our engineers write a scoring set before they touch a prompt, so a change is judged on evidence rather than on whoever tested it last.
- 03If your problem is retrieval or fine-tuning rather than wording, we say so during the free scoping call, before any work starts.
- 04You interview the engineer yourself, and the prompts, scoring sets and pipelines land in your repository from the first commit.
Why "hire a prompt engineer" is usually the wrong request
A prompt engineer edits instructions until an example output looks right. That is real work, and it is also the smallest part of why an AI feature fails in production.
Most inconsistent features fail for one of two reasons. A wording change cannot touch either one. Either the model has no reliable evidence to answer from, or nobody is measuring whether an edit made things better or worse. Greenlight is an ESG research platform. Its fix was pulling 10 to 15 independent sources per topic, so a sustainability score never rested on a company describing itself. PremiumAudit.io is an insurance audit platform. Its fix was reading source documents directly and validating the structured output against them. Calculation errors fell by 95%.
Neither result came from a better-written prompt. Both came from deciding what evidence the model could use and checking the answer against known cases before it shipped.
What we staff instead of a standalone prompt writer
Evals before edits
A scored set of real cases comes first. A prompt change is accepted or rejected against that set, not against how convincing one example looks.
Grounding, not guessing
The model answers from evidence you supply and cite, the way Greenlight cites its sources. A prompt cannot invent facts the system never gave it.
Structured output that validates
Output checked against a schema before your application reads it. Free text parsed by string matching breaks the first time phrasing shifts.
Retrieval when that is the real gap
If answers are wrong because the model cannot find the right document, that is a retrieval problem. No prompt fixes a search that returns the wrong page.
Prompts under version control
Treated as code: reviewed, tested against the scoring set, and rolled back like any other change when a model version updates.
- QuestionFrom a user or a job
- EvidenceWhat the model may use
- PromptInstructions and format
- Structured outputValidated against a schema
- Scored testChecked against known cases
- Ship or refuseLow confidence goes to a person
A standalone prompt engineer usually owns box three. Boxes two, four and five are where inconsistent output actually gets fixed.
Evidence and validation did the work a prompt could not
How hiring works
- 01
Show us the output you do not trust
A free call about the answers that are wrong or inconsistent. If a scoring set and better grounding would fix it faster than a prompt rewrite, we say that first.
- 02
Meet the engineers
We shortlist people who build evals, retrieval and structured output together, not prompt text on its own. You interview them your own way.
- 03
They embed
Your repository, your standups, your review process. One engineer owns the scoring set and can explain every number in it.
- 04
They hand over
Prompts and scoring sets under version control, plus a note on what each model version costs to run. The first 2 months of support and maintenance are free, with every build.
Stack
Models
Orchestration
Practices
Bring us the answer that keeps going wrong
Send an example of the output you cannot trust. The scoping call is free, and you will leave knowing whether this is a prompt problem, an evidence problem, or both.
01Do you have a role called prompt engineer?
No, and we do not think it should stand alone. We staff LLM engineers, who write prompts as one part of a wider job. That job includes building the evidence the model draws from, validating its output, and scoring changes against real cases. Hiring wording help without that structure is how a feature stays inconsistent.
02Our chatbot gives wrong answers. Is that a prompt problem?
Usually not on its own. Check first whether the model had the right information available when it answered. A wrong answer despite good evidence points to the prompt or the model choice. Missing evidence points somewhere else entirely: what the model is allowed to retrieve.
03How is this different from hiring an LLM engineer?
It is not different. An LLM engineer is what a prompt-engineering hire usually turns into once the real problem surfaces. That role adds evals, structured output, retrieval and cost control on top of the prompt itself. See our LLM engineer page for the full role, built on the same case work referenced here.
04Can you just improve our existing prompts without a bigger project?
Yes, when the prompt itself is genuinely the gap and you already have a way to test a change. More often the useful first step is a scoring set, so any prompt edit after that can be judged rather than guessed at. We tell you which situation you are in during the free scoping call.
05What does this cost?
Scoping calls are free. Where we need to look inside an existing codebase to answer honestly, a paid two-week diagnostic replaces the guess with a fixed price.
