PromptGovernance — Internal LLM Prompt Registry
24 production prompts × owner + version + model + eval link + change log + safety review. Per prompt: system message, allowed inputs, banned outputs, kill-switch flag. Surfaces 4 prompts without current eval + 2 orphan prompts inherited from former employee + 3 prompts with safety review overdue + 1 prompt without a kill-switch.
What it is
The shape behind every healthy AI-product team. Most companies build LLM features fast; the prompts accumulate; nobody knows who owns each one; safety reviews lapse; eval coverage drifts. PromptGovernance prototypes the registry that prevents that.
What’s in it
- 24 production prompts across customer support, marketing, sales, dev-tools, BI, growth, retention, internal ops.
- Per-prompt 9-field shape:
- Surface (where it runs — customer support / dev-tools / etc.)
- Model bound + version + last changed
- Owner (or “(orphan)”)
- Eval last run + pass rate (links ModelEvalPanel)
- Safety reviewed at + cadence cap
- Kill-switch wired (feature flag)
- System message (verbatim)
- Allowed inputs / banned outputs explicit
- Notes + cross-references
- Worst-offender findings:
- PG-020 + PG-021 — ORPHAN prompts inherited from ex-employee. Last updated 420 days ago. KILL or take over.
- PG-016 tax-line-explainer — eval AND safety review both overdue (180d).
- PG-007 release-notes-generator + PG-023 dashboard-narrative — safety review overdue.
- PG-014 sentiment-classifier — explicitly notes “do NOT use this to time upsells (see WaitlistEthics EX-015)” — preventing future-team-member from making the same mistake.
- Cross-tool callbacks — every prompt references ModelEvalPanel coverage; the cancellation-flow helper links to WaitlistEthics EX-006 (FTC click-to-cancel); the SQL-translator links to PenTestFindings PT-002; the privacy-questionnaire-prefiller links to TrustCenter SIG.
Why this shape
NIST AI RMF MAP.4 (system governance) + MEASURE.2.6 (ongoing testing) demand the registry. EU AI Act Art 50 (transparency) requires disclosure of AI systems to users. ISO/IEC 42001 §6.2 (AI management system) demands documented prompt + safety governance. OWASP LLM Top-10 demands eval coverage. PromptGovernance prototypes the gate that surfaces orphans + stale evals + missing kill-switches before they cause incidents.
How it ships
Single HTML file, ~19KB. Zero dependencies. 24 prompts × 9-field shape + status classifier in 220 lines of vanilla JavaScript.