KD.

AI lifecycle governance · Independent product work

PromptVault

A lifecycle tool for versioning, structurally checking, and monitoring production LLM prompt assets—with an external evaluation bridge.

Explore the live demo ↗
PromptVault / Versions
Customer Support ReplyVersion history
production
v4

Add empathy instruction

4/4
v3

Shorten to 3 paragraphs

6/8
v2

Initial support tone

8/8
v2 → v3

five paragraphsexactly three paragraphs

The challenge

Prompts often live across documents, configuration, and code without a clear change history, owner, promotion check, or reliable way to investigate what changed.

The outcome

A deployed system connecting immutable prompt history, character-level diffs, structural contracts, owner-scoped access, provider metadata monitoring, and Promptfoo YAML export.

My role: Product definition, causal workflow design, architecture direction, implementation orchestration, claim validation, verification, and deployment.

Selected decisions

Designed around the constraint, not the feature list.

01

Keep versions immutable

Every prompt edit creates a new version so the change history, comparison, and ownership record remain trustworthy.

02

Separate structural and behavioural checks

Zero-cost checks validate rendered prompt contracts; model-output evaluation is delegated to an explicit external workflow.

03

State monitoring boundaries honestly

Model Watch observes provider metadata and availability—not behavioural drift behind an unchanged alias.

Verification evidence

What supports the claim.

  • 19/19 current tests
  • Clean type-check and zero lint errors
  • Owner-scoped PostgreSQL RLS
  • Public 60-second demo
  • Promptfoo export executed end to end

Explicit boundaries

What I do not claim.

  • Structural checks are not model-output regression tests
  • No behavioural-drift detection claim
  • Production email delivery not independently verified
  • No customer, revenue, or traction claim

Continue exploring

Next case study: Chasely →