Claude Code Skills¶
Per jared-zoneraich: in architectural terms, a skill is an extendable system prompt — a lazy-loaded markdown/code bundle that the main claude-code loop opts into when a task matches, rather than stuffing everything into the always-on system prompt.
Purpose¶
Same problem as context-compaction: the permanent system prompt must stay small, but some tasks (docs updates, Word/Excel editing, design style guide, deep research) genuinely need more context. Skills let you ship that context conditionally.
How they're advertised to the model¶
Each skill registers a one-liner description. In theory the model auto-invokes skills whose description matches the current task. In practice, audience member reported Claude ignores their skills and they end up invoking skills manually. Zoneraich agrees this is a known gap: possibly a post-training fix — skill selection needs to become as reliable as tool-call selection.
Case study¶
Zoneraich built the slide deck for this very talk with three skills: slidev (rendering), a deep-research skill, and a design skill — composed via claude-code's main loop.
Tension¶
Flagged as a candidate first-class paradigm that isn't fully cooked yet. Contrast with contextual-prompt-engineering (manual per-task context assembly): skills are the automated version, but auto-selection is unreliable, so many users fall back to manual invocation, which collapses skills back toward slash commands.