I've spent the last 6 months maintaining a 3000-line prompt library for Claude. Yesterday, Anthropic shipped Skills and made 90% of it obsolete. Skills are basically organized folders (SKILL.md files) that Claude loads dynamically when relevant. What makes them interesting:
1. Progressive disclosure - Claude only loads full skill content when needed, not at startup. This scales way better than Custom Instructions or Projects for context management.
2. Composability - Multiple skills stack automatically. Ask it to build a signup form, and it might activate brand-guidelines + accessibility + responsive-design skills without you specifying them.
3. Security model - You can restrict which tools a skill can use via `allowed-tools`. Want a read-only code auditor that literally cannot modify files? That's now possible.
4. Cross-platform - Same skill works in Claude.ai, Claude Code CLI, and the API.
The anthropics/skills repo has some good examples: artifacts-builder (React + Tailwind), webapp-testing (Playwright), document generation, etc.
Real talk though: this isn't "the end of prompt engineering." You're still writing prompts - just packaging them better. But the DX improvement is legit. I went from copy-pasting prompts from Notion to having Claude auto-load the right context based on what i'm doing.
Box, Rakuten, and Canva are apparently already using this in production after 2 weeks, which... seems fast? Rakuten mentioned going from "a day to an hour" on some workflows.
Wrote up my experience here with code examples and the technical details on how progressive disclosure works in the URL.
Current rough edges: no skill marketplace yet, discovery is based on description matching (works surprisingly well but not perfect), sharing is manual (git clone to ~/.claude/skills/), and testing/debugging tooling is basically non-existent.
Has anyone else been playing with this? Curious what workflows people are building skills for.
1. Progressive disclosure - Claude only loads full skill content when needed, not at startup. This scales way better than Custom Instructions or Projects for context management.
2. Composability - Multiple skills stack automatically. Ask it to build a signup form, and it might activate brand-guidelines + accessibility + responsive-design skills without you specifying them.
3. Security model - You can restrict which tools a skill can use via `allowed-tools`. Want a read-only code auditor that literally cannot modify files? That's now possible.
4. Cross-platform - Same skill works in Claude.ai, Claude Code CLI, and the API.
The anthropics/skills repo has some good examples: artifacts-builder (React + Tailwind), webapp-testing (Playwright), document generation, etc.
Real talk though: this isn't "the end of prompt engineering." You're still writing prompts - just packaging them better. But the DX improvement is legit. I went from copy-pasting prompts from Notion to having Claude auto-load the right context based on what i'm doing.
Box, Rakuten, and Canva are apparently already using this in production after 2 weeks, which... seems fast? Rakuten mentioned going from "a day to an hour" on some workflows.
Wrote up my experience here with code examples and the technical details on how progressive disclosure works in the URL.
Current rough edges: no skill marketplace yet, discovery is based on description matching (works surprisingly well but not perfect), sharing is manual (git clone to ~/.claude/skills/), and testing/debugging tooling is basically non-existent.
Has anyone else been playing with this? Curious what workflows people are building skills for.