What's changed: Added per-section figures (cert-figure-retrofit). New AI-300 Chapter 3 (Domain 3 "GenAIOps infrastructure": Foundry environments = project/RBAC/managed identities/private networking/Bicep-CLI; foundation model deployment = serverless API endpoints vs managed compute/use-case model selection/versioning-deployment strategies/PTU; prompt versioning = design-develop/variant comparison/Git)
3.3Prompt versioning and management
Understand designing and developing prompts, creating prompt variants and comparing their performance, and version-controlling prompts with Git repositories—from a source-control and experimentation perspective.
In generative AI apps, prompts are assets as important as code. Because prompt quality strongly affects response quality, you systematically design, compare, and version-control them (operationalizing prompt engineering).
3.3.1Design, variants, and comparison
First, design and develop prompts for the use case (system prompt, instructions, examples, output format). Then create prompt variants (multiple versions with different wording/structure) and compare their performance on the same evaluation data to pick the best. This is the generative-AI experiment process, analogous to "comparing multiple models" in training.
3.3.2Version control with Git
Version-control useful prompts in a Git repository. This enables change history, review (pull requests), and rollback, and lets you trace "which prompt version produced which evaluation result." Managing code, IaC, and prompts all in Git makes GenAIOps reproducible and auditable.
| Stage | What | Key point |
|---|---|---|
| Design/develop | System prompt/instructions/examples/output format | Fit the use case |
| Variants | Multiple versions, different wording/structure | Compare on the same data |
| Versioning | Git repository | History/review/rollback |
Watch the mix-ups: (1) Prompts are versioned assets, not disposable. (2) Compare variants on the same evaluation data (control conditions). (3) Manage code, IaC, and prompts in Git to make GenAIOps reproducible.
Map requirement → means: "compare multiple prompts with different wording" = prompt variants + same evaluation data; "change history and review of prompts" = version control in Git; "trace which prompt version gave which result" = Git + linking to evaluation results.
3.3.3Section summary
- Prompts are assets: design/develop for the use case
- Compare variants on the same evaluation data and pick the best
- Version-control in Git (history/review/rollback) to make GenAIOps reproducible
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. In generative AI apps, what should be managed systematically as an asset because it strongly affects response quality?
Q2. To pick the best among multiple prompts with different wording/structure, what should you do?
Q3. Which enables change history, review, and rollback of prompts?
Q4. Which is NOT a typical element specified when designing/developing prompts?
Q5. Which consistent approach makes GenAIOps reproducible and auditable?
Q6. Why is prompt versioning analogous to "comparing multiple models" in training?