Concept card

Prompt Libraries

As of 2026-10-04

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is Prompt Libraries?

Senior consultants run 50-200 versioned prompts as reusable assets, refactored every quarter as models evolve — treating prompts like code, not throwaway text.

What it is

A prompt library is a maintained, versioned collection of prompts that a team reuses — with each entry owned, tested against known inputs, and retired when the model or the task changes.

Why it matters

The unmanaged alternative is every consultant keeping private prompts in a notes app, which produces three costs: the same problem is solved repeatedly, quality varies invisibly between people, and nothing improves because no one sees anyone else's failures.

The library is also where prompt governance becomes possible. A prompt that instructs a model to summarise client data is a data-handling decision, and it is auditable only if it exists somewhere findable.

How it works

Treat each prompt as a small product: a name, the task it solves, the model it was tuned against, an example input and its expected shape of output, and a date. The date matters more than it looks — prompts are coupled to model versions, and one that worked reliably against a previous generation can degrade silently.

Test with the boring cases, not the impressive ones. A prompt is validated by behaving predictably on ordinary input, not by producing one striking result in a demo.

The senior move

Version the prompt with its evaluation, not just its text. Storing what good output looked like at the time is the only way to detect that a prompt has drifted when the underlying model changed under it — otherwise the degradation is discovered by a client.

Why it matters in practice

A prompt written once and never revisited degrades as models change — quarterly refactoring is the maintenance cost of staying current.
50-200 prompts is the working range for seniors; fewer signals ad hoc use, not a real asset base.

Key points

A versioned, organized collection of high-performing prompts treated as reusable assets.
Senior consultants run 50-200 prompts, refactored quarterly as models evolve.
A prompt is versioned with its evaluation, not just its text — storing the expected output shape is what lets drift be detected when the underlying model changes.
SAP's Prompt Registry (generative AI hub) is the enterprise-scale version of this discipline: centrally governed, versioned templates with a Git-based sync API for CI/CD.
The orchestration service's templating module is where a registered prompt is actually used inside a pipeline — separating prompt text from pipeline configuration (grounding, filtering, masking) rather than hardcoding one string.
A prompt library with no owner and no version history is invisible to AI Agent Hub-style governance sweeps — 'which prompts touch personal data, and who approved them' has no answer without one.
Quarterly maintenance means re-running stored example inputs against the current model and diffing the output shape — not rewriting the prompt speculatively.

Terms used on this page

Judgment layer: The part of the work that AI cannot do — prioritisation, trade-offs, client-context reading.
AI-first draft: Workflow where AI produces the first pass (code, memo, slide) and the consultant edits rather than writes from scratch.
Agent reliability: The consistency, cost, safety, and policy compliance of an agent across repeated runs.
Prompt Registry: SAP's centralised, versioned store for prompt and orchestration-config templates in generative AI hub, synced via a declarative Git-based API for CI/CD.
Golden set: The stored example inputs and their expected output shape against which a prompt is re-tested after a model or task change — the artifact that turns 'it felt different' into a measurable diff.
Templating module: The orchestration-service pipeline step that fills a registered prompt's placeholders at run time and composes it with grounding, filtering and masking modules.
Prompt drift: The silent degradation of a previously reliable prompt's output quality after the underlying model changes generation — the reason prompts are versioned with a date, not treated as permanent.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →