What problem does it solve?
Onboarding a new image or video generation ecosystem (e.g. a new Flux variant or Wan video version) requires writing a tuned system prompt for the prompt-analysis service, and doing it ad hoc produces guides that drift from the model's real behavior or saturate the analyzer with repetitive recommendations.
Core Features & Use Cases
- Guide authoring workflow: Researches the ecosystem from a model card or spec, maps findings to a fixed guide template, and enforces measured tone rules (no absence-check guidelines, no emphatic capability claims, no out-of-payload conditions).
- Measurement and audit tooling:
measure.mjs A/B-tests a candidate guide against the live one using topic-saturation metrics, and audit.mjs statically screens every live guide for known defect patterns.
- Safe deployment:
manage.mjs registers, updates, exports, and rolls back guides on the orchestrator's /v1/manager/prompt-analysis endpoints with write gating, readback polling, and sample preservation.
- Use Case: When a new video model like MiniMax H3 ships, provide its ecosystem key and HuggingFace model card, and the skill drafts a guide consistent with sibling ecosystems, measures it against the live baseline, and deploys it with few-shot samples.
Quick Start
Ask the assistant to add a prompt-enhancement guide for a new ecosystem, providing the ecosystem key from basemodel.constants.ts and a link to the model card or prompting documentation.
