Everything You Know About Skills IS OUTDATED

This video reviews Anthropic’s updated best practices for building Claude skills, explaining new rules for reference files, degrees of freedom, model-specific testing, file nesting, checklists, self-validation loops, and dependency handling. The presenter walks through practical fixes—contents lists, scripts, one-level references—and how to make skills portable, robust, and maintainable.

– Contents lists: Add an index to any reference file over 100 lines so Claude can jump directly to needed sections.
– Degrees of freedom: Apply high/medium/low freedom per step; use scripts for low-freedom, fragile operations (payments, deletions).
– Model testing & structure: Test skills across target models, note model-specific detail levels, keep references one level deep and skill.md under 500 lines.
– Validation & portability: Use checklists and self-feedback loops to iterate outputs; include explicit install commands for dependencies so skills work on fresh environments.

Quotes:

If your important rules sit after line 100, as far as Claude’s concerned, they actually don’t exist.

One skill can mix freedom levels: creative writing and a locked-down script in the same workflow.

Scripts run instead of being loaded, so they don’t consume Claude’s context.