The 5 Claude Code skills
that actually fire.
And the configs
that make them run.
You commented FIRE. Here it is. The skills are the easy part. The invocation is what nobody talks about.
Quick truth before we get to the list. Claude Code skills are model-invoked. That means they only fire when your context tells Claude to load them. Most people install 50 and use 0 because their CLAUDE.md, project structure, and prompts never wire the trigger.
The 5 below are the only ones I actually run every day. For each one I'll show you the install path, then the why most people skip it note that explains the failure mode, then the //try: command that actually works on day one.
The invocation map at the bottom is the part Marc never published.
Graphify / 71.5x less tokens
Turns any folder of code or docs into a queryable knowledge graph. Instead of stuffing 80,000 tokens of context into every prompt, Claude queries the graph and pulls only what it needs. The 71.5x token reduction is real on large codebases. The infinite memory framing is real on long-running projects.
UI/UX Pro Max / $10K sites
50+ design styles, 161 palettes, 57 font pairings, all callable as a system inside Claude Code. Pair it with the 21st.dev MCP and you get studio-grade animated components on demand. This is the skill that stops Claude from defaulting to the AI-slop look (Inter, purple gradient, tile cards).
/ui-ux-pro-max once at the start of a project and assume it's loaded forever. It isn't. You re-invoke it every time you start a new component, or you anchor it in CLAUDE.md so the design system loads on every relevant turn. Most people do neither, then complain Claude's UI looks generic.
Claude SEO / Auto-Fix
Runs 19 sub-audits across technical SEO, schema, meta tags, page speed, AI search readiness, and E-E-A-T signals. Then it implements the fixes inside your repo. Not a report. The actual changes, committed.
/seo-audit once and never set up the agent loop. The skill is designed to run as a subagent that watches your sitemap and reruns on every deploy. Most people treat it as a one-shot. The compounding value lives in the loop.
Remotion / Claude edits videos
Describe your video in plain english. Claude builds the React composition. Animation timing, captions, audio sync, data viz, all generated. 6M+ views on launch. 117K+ weekly installs as of April 2026.
remotion-best-practices companion skill. Claude without best-practices writes Remotion code that compiles but renders broken. The pair is non-negotiable. Install both or don't bother.
Google Workspace CLI / gws
Claude controls Gmail, Drive, Docs, Sheets, Calendar, Slides, Chat, and Admin through a single unified interface. 100+ built-in actions. Ships with an MCP server. 4,900 GitHub stars in the first 3 days because nobody had wired Workspace into Claude this cleanly before.
gws mcp with all APIs enabled. That blows out the context window with permission scopes Claude doesn't need. Run it with -s drive,gmail,calendar only. Specific is faster. Specific is also what stops Claude from accidentally writing to the wrong inbox.
The Invocation Map
Skills only fire when context triggers them. Here's the four-line check that decides whether ANY skill in your .claude/skills/ directory will actually run today.
"For any frontend task, invoke /ui-ux-pro-max first."
/seo-audit beats hoping Claude figures it out from a vague prompt.
Run all four. Watch your hit rate go from 10% to 90%+ on the same skills you already had installed.