What problem does it solve? When building or refactoring AI agent skills, it is hard to know what belongs in SKILL.md versus references/ or scripts/, and bad splits waste context tokens, break triggering assumptions, or hide critical content from the model. ## Core Features & Use Cases - Split decision framework: Separates triggering problems (fix the YAML description) from disclosure problems (file splitting), and defaults to keeping skills monolithic until evidence justifies a split. - Three proven split axes: Guides variant-branch splits (e.g., AWS/GCP/Azure), workflow-vs-reference-data splits (procedures vs. lookup tables), and depth-tier splits (common path vs. edge cases), each with explicit pros and cons. - Anti-pattern detection and architecture evals: Flags topic-based splits, cosmetic splits, and rare-but-critical content in references/, plus provides instrumentation, load-rate metrics, and decision rules (e.g., merge references co-loaded over 70% of runs). - Use Case: Your SKILL.md has grown to 600 lines covering three cloud providers. This skill recommends a variant-branch split with one reference file per provider and writes the exact pointer sentences for SKILL.md. ## Quick Start Ask the skill to review your SKILL.md and recommend whether and how to split its content into reference files.
