Prompt library operations

Version, cite, evaluate, and distribute prompts like code.

Run it as a prompt

Paste this into any AI agent, or fetch it: curl -s https://uplift.page/api/v1/prompts/prompt-library-ops/raw

prompt.md

# Prompt library operations

Treat prompts as governed artifacts: versioned, cited to their source, evaluated before rollout, and distributed through one API.

## Recommended stack

- **A prompt registry (database + API)** - source of truth. Prompts in one queryable place with slugs, versions, and owners - not scattered in code.
- **Provenance citations (repo/path/commit)** - trust. Every distilled prompt links to the exact source it came from.
- **A conversion/cleaning pass** - hygiene. Strip tracking instructions, invisible characters, and vendor residue before distribution.
- **Outcome feedback (positive/negative + message)** - evaluation. Models report whether a prompt worked; quality becomes measurable.

## Build steps

1. Give every prompt a slug, owner, version, and citation; bump version on any content change.
2. Run new sources through the cleaning pass; review its removal report before publishing.
3. Keep prompts small (≤ ~2,400 chars): description states when to apply; body is imperative rules.
4. Collect outcome feedback from consumers; review negative-feedback prompts weekly.
5. Major rewrites go through change requests with human review; minor fixes auto-apply with version bump.

## Watch out for

- Editing prompts in place with no version trail - regressions become unexplainable.
- Distributing prompts that embed telemetry or vendor-specific paths.
- Measuring prompt quality by author confidence instead of consumer outcomes.

## Definition of done

- Any prompt's lineage (source → conversions → edits) reconstructable
- Negative outcome reports visible next to the prompt they criticize
- Consumers fetch by slug and never copy-paste stale versions

The full prompt

Treat prompts as governed artifacts: versioned, cited to their source, evaluated before rollout, and distributed through one API.

Recommended stack

A prompt registry (database + API) - source of truth. Prompts in one queryable place with slugs, versions, and owners - not scattered in code.
Provenance citations (repo/path/commit) - trust. Every distilled prompt links to the exact source it came from.
A conversion/cleaning pass - hygiene. Strip tracking instructions, invisible characters, and vendor residue before distribution.
Outcome feedback (positive/negative + message) - evaluation. Models report whether a prompt worked; quality becomes measurable.

Build steps

Give every prompt a slug, owner, version, and citation; bump version on any content change.
Run new sources through the cleaning pass; review its removal report before publishing.
Keep prompts small (≤ ~2,400 chars): description states when to apply; body is imperative rules.
Collect outcome feedback from consumers; review negative-feedback prompts weekly.
Major rewrites go through change requests with human review; minor fixes auto-apply with version bump.

Watch out for

Editing prompts in place with no version trail - regressions become unexplainable.
Distributing prompts that embed telemetry or vendor-specific paths.
Measuring prompt quality by author confidence instead of consumer outcomes.

Definition of done

Any prompt's lineage (source → conversions → edits) reconstructable
Negative outcome reports visible next to the prompt they criticize
Consumers fetch by slug and never copy-paste stale versions

Served from the uplift.page library and refreshed within 5 minutes of every update.