maintaining-ai-models
Installation
SKILL.md
Maintaining AI Models
A model choice goes stale without anything breaking: the provider withdraws an ID (and may redirect it to a "comparable" model, a silent behaviour change), announces a shutdown, or ships a better model nobody measures. This skill keeps that question answered with two tiers:
| Tier | What it answers | Cost | Tool |
|---|---|---|---|
| Check (deterministic) | Is every model still offered? Is a shutdown announced? Does the code match the adapter? Does the model still do what the usage needs? | No Claude tokens; the probes cost a few provider tokens | ai-models.mjs check [--probe] |
| Measurement | Which model does the job best in this application? | A full benchmark run | /lt-dev:backend:ai-model-review |
Quality is measured only through the project's real pipeline, never with raw API calls: prompts, batching, quality gates and parsers decide together with the model. The check only says whether that run is needed.