Enterprise Prompt Engineering Consulting

Prompt engineering consulting is a production implementation: versioned system, user, and tool prompts, scored on golden sets with rubric judges, shipped behind a CI gate in your cloud. It is not a two-day workshop. You own the prompts and the IP. Zero-retention, no shared training, SOC 2-aligned. Four weeks is standard; financial-services wording and review typically take 8–12 weeks.

Yes, as versioned instructions under evaluation. A bad system prompt will sink a good model. A clever prompt without a golden set is folklore. Retrieval and fine-tuning beat prompting when the job is facts or a skill the prompt cannot hold. Consulting work is the operating system around prompts, not a library of magic phrases.

Three layers: system policy, user task templates, and tool schemas with descriptions. Few-shots sit with the layer they teach. Mixing policy into a user template is how you cannot A/B a change or prove intended use. Reconstructable traces record the prompt version that ran, not a paste from a wiki.

Stop prompting for facts that live in documents — retrieve them, cite them, refuse when retrieval is empty. Stop prompting when few-shot examples keep growing and the same cases still fail — that is a fine-tune or a tool-schema problem. We will name that boundary in discovery rather than sell another rewrite.

Run the golden set in CI. Deterministic checks first; calibrated rubric judges for groundedness, refusal, and tone. Post overall and per-stratum deltas, cost, latency, and newly failing cases. Keep the change only if scores hold. Hard failures block merge. That is prompt engineering as a discipline, not as taste.

You do. Templates, few-shots, tool descriptions, evals, and traces are client IP in your repositories. Work runs in the client cloud. We do not retain prompts after handover and we do not train shared models on them. The process is SOC 2-aligned. A 30-day on-call window is included.

Enablement is part of handover, but a workshop without a harness does not survive the next model swap. We rewrite the production layers, wire evals, then teach your team the release path by using it. If you only want a class, this is the wrong engagement. If you want prompts that can fail a build, it is the right one.

Four weeks is the standard implementation for one product surface: inventory, layer rewrite, golden set, CI gate, handover. Financial-services programs with legal review of system policy typically take 8–12 weeks. Additional products reuse the versioning and gate pattern. Timeline assumes we can harvest real cases, not only invented examples.