Anthropic Agent Skills
Claude Agent Skills Explained is a compact primary-source video about a practical layer in agent design: packaging reusable expertise so an agent can load the right procedure, examples, scripts, or reference material only when the task calls for it. The video explains skills as portable folders, then separates them from three adjacent mechanisms: project memory in CLAUDE.md, MCP connections to outside data and tools, and subagents with their own role, context window, and permissions.
The strongest Spiralist relevance is institutional habit becoming machine-readable. A skill can encode coding standards, security review practice, data-analysis methods, accessibility expectations, document workflows, and other local norms that used to live in people, onboarding, and review culture. That belongs beside AI Agents, Model Context Protocol, AI Coding Agents, Agent Tool Permission Protocol, and Agent Audit and Incident Review. The governance problem is not only whether Claude can follow a skill; it is who authors the reusable procedure, who reviews it, what tools it can invoke, and whether the resulting work leaves a trace humans can inspect.
Anthropic's Agent Skills documentation supports the video's core architecture: skills are filesystem-based resources with metadata, instructions, and optional scripts or references, loaded when relevant rather than placed entirely into context at startup. Anthropic's Skills comparison guide and subagents documentation support the distinction between reusable expertise, project context, MCP integrations, and independent delegated agents. Anthropic's engineering post on Agent Skills gives the broader rationale: agents need specialized work knowledge, not only stronger base models.
Uncertainty should stay visible. The video is an official product explainer, not an independent evaluation of skill reliability, permission boundaries, or enterprise governance. It does not prove that skills eliminate context failure, prompt injection, stale procedures, unsafe scripts, or weak human review. Treat it as strong evidence of Anthropic's late-2025 agent architecture: reusable procedural memory is becoming a first-class surface for delegated work.