A practical guide covering the top five pitfalls when writing and reviewing Agent Skills for agentic coding systems like Claude Code. Key advice: skill descriptions should only convey when to trigger the skill (30-55 words, imperative); use progressive disclosure by pushing scenario-specific content to reference files; offload deterministic/repetitive work to scripts (preferably Python with uv); keep skills lean and apply a deletion test; and structure multi-step workflows with numbered steps and a final review. Also clarifies the distinction between Custom Agents (persona-based, own context) and Skills (composable, portable, knowledge/workflow-focused).
316 Impressions