Versioning prompts

Here is the mistake that ends most AI-feature incidents in confusion: the prompt lives in a chat window, a Notion doc, or an un-versioned string, and nobody can say which prompt produced the bad output. You've kept version numbers by hand since Prompt Engineering; LLMOps makes prompt versioning real infrastructure, because in production the prompt is the code — and code you can't version, you can't operate.

What a 'version' is

A prompt version is not just the prompt text. It's the whole configuration that determines behavior — because changing any of it changes the output:

• The prompt template(s) — system and user.
• The examples (few-shot content).
• The model + settings (claude-sonnet-5, max_tokens).
• Retrieval/tool config, if any (k, which tools, thresholds).
• A version id and a change log entry: what changed, why, and its eval delta.

VERSION = "triage-v7"
MODEL = "claude-sonnet-5"
MAX_TOKENS = 500
SYSTEM = """...full system prompt..."""
EXAMPLES = [ ... ]
# CHANGELOG (newest first)
# v7: added forwarded-message boundary example.
#     eval: golden-v3 pass 88% -> 93%; holdout 90%. Ships.
# v6: tightened urgency definitions. 84% -> 88%.

Prompts in source control

• Prompts live in the repo, not the app UI. A change to a prompt is a pull request — diffable, reviewable, revertible. A changed gold answer or a reworded rule gets the same scrutiny as changed code, because it has the same blast radius.
• Every deployed output is stamped with its version. The trace's version field (Module 3) plus the versioned prompt means 'which prompt said that, on which model?' is a lookup, not an archaeology project.
• Version = eval score = traces. The three link into one record: this version scored 93% on golden-v3, and here are its production traces. That linkage is what makes the loop auditable — you can prove what shipped, what it scored, and how it behaved.