Thoughts on Claude Skills (bit.kevinslin.com)

Claude Skills are a new way to package LLM behavior as lightweight, structured markdown bundles: a SKILL.md with simple frontmatter (name, description) and optional supporting files (docs, scripts, assets). At runtime the model only reads the title/description up front (saving tokens) but can access the full skill content and bundled code when invoked, letting you ship everything from emoji catalogs to utility scripts (e.g., hello/SKILL.md, EMOJIS.md, scripts/animate.py). Skills combine the simplicity of AGENTS.md-style prompts with the power of MCP/tool integration, but without forcing the model to load long function signatures into every conversation. For practitioners this is a practical inflection point: skills address context bloat and scale problems that AGENTS.md and MCPs encountered, enabling composable, reusable “objects” of LLM behavior that users can assemble without retraining models. That improves token efficiency, workflow portability, and local extensibility (author built a CLI “skillz” to inject skills into other tools). Current limits: skills are Claude-only today and lack native sharing/distribution mechanisms. Still, they hint at a future where richer ecosystems, better sharing, and higher-level abstractions let developers extend model capability by composition rather than waiting for larger models. If you’re hitting context limits or copy-paste prompt debt, skills are worth exploring.