Skip to main content

Skills

Skills are Claude Code's on-demand knowledge system. Each skill is a directory containing a SKILL.md file with YAML frontmatter and markdown instructions — loaded into context only when relevant, keeping your baseline token budget clean.

Why Skills Matter

CLAUDE.md burns tokens on every turn whether the instructions are relevant or not. Skills solve this with progressive disclosure: Claude sees a lightweight registry of skill names and descriptions (~100 tokens each), then loads the full body only when the task demands it. Twenty skills cost ~2,000 tokens of always-present overhead. The same content in CLAUDE.md would cost that on every single turn regardless of relevance.

When to Use Skills vs. Alternatives

NeedUse
Task-specific multi-step workflowSkill
Always-present project rulesCLAUDE.md
Deterministic enforcement of constraintsHook
Block dangerous operations before executionHook (PreToolUse)
Reusable procedure across multiple reposSkill (personal ~/.claude/skills/)
One-off saved promptLegacy command (.claude/commands/)
External tool connectivityMCP server
Distribute workflows as packagesSkill (plugin)

The critical distinction: skills extend what Claude can do on demand. Hooks constrain how Claude does it deterministically. CLAUDE.md tells Claude what it should do at all times.

What's Inside

Explore the sub-pages for deep expertise:

  • Mental Model — How skills are loaded, matched, and injected through the three-stage progressive disclosure system
  • Playbook — Production skill architectures: knowledge-based, workflow, checklist-driven, and forked execution patterns
  • Compositions — Skills combined with hooks, agents, MCP, CLAUDE.md, and memory for orchestrated workflows
  • Pitfalls — Activation failures, context bloat, conflicting skills, and the anti-patterns that make skills unreliable