Systems on Glass is an AI-automation side project of mine. Every episode is generated by Claude (Anthropic's AI, working through Claude Code in my own video studio): the script, the animation and 3D scenes, the edit, and the narration in an AI clone of my voice. I choose the topics, bring the stories from my own work, and review each episode before it goes up. Not affiliated with Anthropic.
The agent that makes this series knows how to make every video I make: the styles, the voice, the checks, the publishing. That's 205,859 characters of how-to in 34 files. When a session starts, it reads 893 of them.
The short version
- A skill is a folder. A
SKILL.mdwith a name, a description and the steps, plus whatever the steps need: guides, references, scripts. Anthropic's engineers compare it to an onboarding guide for a new hire. - Level one, the label. At startup the agent loads only the name and description of every installed skill. For my studio, 893 characters. Unlike
CLAUDE.md(episode 9), a skill's body loads only when it's used. - Level two, the steps; level three, the tools.
SKILL.mdloads when a task fits (mine: check setup, interview me, research what's current, offer three styles, estimate first). Each step points to files it opens only when needed, and scripts run without being read into context. - The label is the trigger. The agent only sees the description, so write it with the words you'd actually say ("I want to create a video", "make a reel").
- It keeps learning, and it travels. My skill has 48 commits since October; today's added "research what's current". Since December 2025 skills are an open standard (agentskills.io). The catch: a skill is code you let in, so install only from trusted sources.
"make a video": + SKILL.md, 19,353 · a step needs a guide: + 11,222 · the rest waits
The numbers come from the skill that made this episode, measured on November 7, 2026.
Further reading
- Equipping agents for the real world with Agent Skills — Anthropic's engineers on progressive disclosure, scripts and security (October 2025).
- Agent Skills, the open standard — the format any agent can read (December 2025).
- Extend Claude with skills — the Claude Code docs: where skills live, frontmatter, and how they differ from CLAUDE.md.
Next episode: Idempotency. Previously: Retries, Backoff and Jitter. All episodes →