What problem does it solve? Maintaining accurate knowledgeCutoff, family, and generation metadata across roughly 1900 model entries in 80 provider files is error-prone: vendors publish cutoffs inconsistently, aggregator sites copy wrong values, and naming conventions hide family and generation information. This Skill encodes the sourcing rules, derivation regexes, and codemods needed to fill these fields correctly at single-model or repo-wide scale. ## Core Features & Use Cases - Sourcing policy for knowledgeCutoff: Accepts only vendor docs, official Hugging Face cards, and tech reports; rejects aggregator sites and AWS Bedrock cards that conflate launch dates with cutoffs. - Rule-based family/generation derivation: scripts/derive-family.ts maps normalized model ids to family and generation slugs, handling traps like date suffixes, size suffixes, and vendor spelling variants. - Idempotent codemods: scripts/apply-cutoffs.ts and scripts/apply-family.ts insert fields into prettier-formatted model card files, skipping entries that already have values. - Use Case: When a new Claude model PR lands, fill all three metadata fields inline citing the official system card, or run a repo-wide sweep with multi-agent research plus adversarial verification to backfill cutoffs across all providers. ## Quick Start Ask the AI to derive and apply family and generation metadata for all chat models in packages/model-bank/src/aiModels using the skill's extract, derive, and apply scripts.
