Prompt Engineering for SAP

As of 2026-05-29

Prompt engineering for SAP is narrower than generic prompting: the goal is a grounded, trustworthy, governed answer about the client's SAP data — the skill is in grounding, guardrails, and structure, not clever wording. Grounding is ~90% of the work: RAG over governed data products + semantic layer (M033) before the model answers — what makes Joule (M046) trustworthy and a raw LLM dangerous. Structure beats cleverness: clear role/task + grounded context + structured output + guardrails ("answer only from provided context; don't invent; cite"). The key anti-hallucination instruction enforces §0c grounded-AI + §0i no-bullshit. Prompts live in inline Joule, custom Joule Studio skills, and agent instructions (M049). Evaluate prompts like code (version/test/review). Honest caveat: grounding reduces but doesn't eliminate failure — promise "grounded, cited, evaluated", not "always right".