Workshop: Skills vs Instructions

Question

When should a reusable procedure be a skill and when is it enough to add the path to an instruction file in CLAUDE.md?

Both are distilled procedures in the same medium. Both need a name and description always loaded. Both load the body on demand. The agent follows both autonomously. The file format is interchangeable. So what, if anything, actually differs in the loading mechanism — and does that difference matter for when to use which?

Grounding theory

• Skills are instructions plus routing and execution policy — revised claim (was "instructions are skills without automatic routing")
• Skills derive from methodology through distillation — the shared derivation process
• Instruction specificity should match loading frequency — the loading hierarchy both participate in
• Always-loaded context mechanisms — the surfaces through which both are discovered
• Agent memory needs discoverable, composable, trusted knowledge under bounded context — evaluation criteria (discoverable, composable, trustworthy)

Research plan

See search-plan.md.

---

Complete file listing (generated at build time)