Skills vs prompts vs a prompt library: what goes where

Skills answer "HOW do we do this procedure?" An Agent Skill is a folder — a SKILL.mdwith instructions, optionally scripts and resources — that the agent discovers and loads on demand when a task matches. Simon Willison called Skills maybe "a bigger deal than MCP," and the design reason is token economics: only the skill's name and description ride in standing context; the full procedure loads when needed. Skills shine for multi-step procedures with tooling: "how we generate PDFs," "how we run migrations."

Rules files answer "what is ALWAYS true here?" CLAUDE.md and its cousins are standing context — loaded every session, whether relevant or not. That's exactly why they must stay lean (the 300-token rule): decisions only, no procedures, no task material.

A prompt library answers "what do I SAY for this task?" Task prompts — review this diff, plan this feature, audit this boundary — are invoked deliberately, vary by situation via variables, and improve with use. They don't belong in standing context (attention tax) or in a Skill (they're judgment-shaped, not procedure-shaped): they belong somewhere versioned, searchable, and one command away from any surface.

Is it true in EVERY session?              → rules file (keep it tiny)
Is it a PROCEDURE with steps/scripts?     → Skill (loads on demand)
Is it a TASK you invoke and refine?       → prompt library (versioned)

Litmus: if you'd want a diff when it changes and a rollback when
it regresses, it's library material.

The three compose instead of competing: a lean rules file names the constitution, Skills package the procedures, and the library holds the judgment prompts both of them invoke. An agent mid-Skill can even pull a library prompt over MCP — the layers stack.

Content drifts between layers as it matures, and the drift is the point. A one-off task prompt grows a checklist, the checklist stabilizes into procedure, the procedure becomes a Skill. Or the reverse: a bloated rules file sheds its task material into the library and drops from 3,000 tokens to 400. Watching version historytells you when a prompt is ready to graduate — and keeping the layers explicit is what makes the moves deliberate instead of accidental. Our own split is public: standing rules stay in each project's CLAUDE.md; everything task-shaped lives in the library, one ple get away.