2025/Free online book · Field Manuals
From Prompt to Pipeline
Turning a clever prompt into a system you can operate
Access
Free
Chapters
12
Read time
120 min
Alpesh Nakrani
VP of Growth at ViitorCloud · former CTO & COO
Every AI feature starts as a prompt that worked once. This manual is the unglamorous middle: versioning prompts, handling the long tail of inputs, building the observability that turns a clever result into a system someone can run on call.
A prompt that works in the playground is not a product. The path from a good result to a system that holds at 3am.
This edition is free to read onsite. Each chapter has its own URL, so readers can bookmark, share, and return to the exact section they need.
Table of contents
INTIntroduction: When a Demo Prompt Started Acting Like a ProductA prompt that answers well in a demo is mistaken for a system, and the gap between the two decides whether the feature survives contact with real customers.10 min01The Long Tail Waiting Behind the DemoThe demo shows the ten inputs someone chose. Production shows the ten thousand inputs nobody chose, and the prompt meets all of them cold.9 min02A Prompt Registry Gives the String an OwnerA prompt with no version number is a rumor. A prompt registry turns it into an asset someone can be accountable for.9 min03Input Policy: What You Accept Before the Model Sees ItEvery prompt inherits whatever the user typed, unless something upstream decides what counts as a legitimate request first.10 min04Output Contracts: What You Accept Before the Customer Sees ItA model's answer is a draft until something checks it against a contract, and most teams ship the draft.8 min05Fixtures and the Regression Suite a Prompt NeedsA prompt change that reads better on one example can quietly break ten others, and only a fixture suite catches it before a customer does.8 min06Review Is Not Optional for a Customer-Facing PromptA one-line prompt edit can change refund policy, medical caution, or legal exposure, and it deserves the same review a line of code gets.8 min07Rollback and the Release Note Nobody WroteThe fastest way to fix a bad prompt is to already know what the last good one looked like, and most teams find that out the hard way.8 min08Observability for a System Made of SentencesA prompt-driven feature fails in ways a stack trace cannot show you, and you need a different instrument to see it happen.9 min09The Model Upgrade You Did Not ChooseA vendor's silent model swap can rewrite your feature's behavior overnight, and the invoice will not tell you it happened.8 min10Where a Prompt Chain Becomes a PipelineTwo prompts stitched together for a demo become five in production, and at some point the stitching itself becomes the thing you have to operate.8 min11Ownership and the On-Call Rotation for a PromptIf nobody owns a prompt at 3am, nobody owns it at 2pm either, and the feature is one bad output away from finding that out.8 min12The Day an Experiment Becomes a WorkflowAn experiment is allowed to break. The day customers depend on it, that permission quietly expires, and someone has to notice.9 minENDConclusion: Prompts Are Operational Assets, Not Lucky StringsThe habit that closes this book: give every prompt a version, an owner, a test suite, and a rollback plan before it becomes load-bearing.8 min