PM-201c · Module 2

Prompt Version Control

3 min read

"I updated the prompt" is a statement that, without version control, carries no information about what changed, why, or whether the change improved or degraded the output. It is a declaration that the past no longer exists. In a prompt system that affects production workflows, this is a liability. Version control for prompts is not optional. It is the minimum viable governance that makes a prompt system auditable, reversible, and improvable.

Semantic versioning for prompts:
MAJOR.MINOR.PATCH

MAJOR: Changes to the output schema, the fundamental task, or the role definition.
  Example: Adding a new required output field → 2.0.0
  Example: Changing from summarize to classify → 2.0.0
  Requires: new golden dataset, full regression test, team notification

MINOR: New constraints, tone changes, context additions that change output behavior.
  Example: Adding a negative constraint for pricing → 1.2.0
  Example: Adding tone specification → 1.2.0
  Requires: test against existing golden dataset, document in changelog

PATCH: Wording fixes that do not change output behavior.
  Example: Fixing a grammatical error in the role clause → 1.1.1
  Example: Reordering constraints for clarity → 1.1.1
  Requires: spot-check, document in changelog

Changelog format:
[version] [date] [author] — [what changed] — [why changed] — [tested against]
Example:
3.0.0 | 2026-03-01 | forge | Added "must not include pricing" negative constraint |
  Feb incident: model included approximate pricing when client context was provided |
  Tested against golden dataset v2 (15 samples, 100% compliance)