Every figure, name and mechanism the finished picture puts on screen, chased to a primary source. The video is largely an argument about how to work rather than a claim about numbers, so the checkable surface is small and almost all of it is the definition of a Skill itself.
A Skill is a directory containing a SKILL.md file, and SKILL.md must open with YAML
frontmatter carrying name and description.
“Every Skill requires a
SKILL.mdfile with YAML frontmatter … Required fields:nameanddescription.”
name is capped at 64 characters and must be lowercase letters, numbers and hyphens;
description is capped at 1024 characters and must state both what the Skill does and when
Claude should use it.
Source: Anthropic, Agent Skills, https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
A Skill directory may hold more than SKILL.md: additional markdown instructions,
executable scripts, and reference resources.
“Instructions: Additional markdown files … Code: Executable scripts … that Claude runs using bash … Resources: Reference materials such as database schemas, API documentation, templates, or examples”
Source: Anthropic, Agent Skills, as above.
The three loading levels the picture puts on screen in 021-loads-when-needed, with the
figures Anthropic publishes for each:
| Level | When it loads | What it costs |
|---|---|---|
name and description from the frontmatter |
always, at startup | about 100 tokens per Skill |
the SKILL.md body |
when the Skill is triggered | under 5k tokens |
| bundled files and scripts | only when read or run | nothing until accessed |
“Claude loads this metadata at startup and includes it in the system prompt … until a Skill is triggered, only its name and description occupy context.”
“Progressive disclosure ensures only relevant content occupies the context window at any given time.”
Source: Anthropic, Agent Skills, as above (the “Level / When loaded / Token cost” table).
These are the only three figures the video renders as numbers, and all three come from that table.
Filesystem based, and not uploaded: ~/.claude/skills/ for personal Skills,
.claude/skills/ for project Skills. The picture uses project relative paths such as
skills/repo-map/SKILL.md throughout, which matches that arrangement.
Source: Anthropic, Agent Skills, “Claude Code” section, as above.
Claude Code installs plugins from marketplaces, and a plugin can bundle Skills, agents, hooks and MCP servers.
“build one or more plugins with skills, agents, hooks, MCP servers, or LSP servers”
A marketplace is added with /plugin marketplace add, which is the command shown on screen
in 181-plugins-can-include-skills.
“users add your marketplace with
/plugin marketplace addand install individual plugins”
Source: Anthropic, Create and distribute a plugin marketplace, https://code.claude.com/docs/en/plugin-marketplaces
Plugins were announced as a public beta covering slash commands, subagents, MCP servers and hooks.
Source: Anthropic, Customize Claude Code with plugins, https://www.anthropic.com/news/claude-code-plugins
Skills are read off the filesystem with bash when triggered, and a bundled script’s code
never enters the context window at all; only its output does. That is what
019-scripts-templates-examples and 199-reusable-instructions-on-demand draw.
“When Claude runs
validate_form.py, the script’s code never loads into the context window. Only its output … consumes tokens.”
Source: Anthropic, Agent Skills, as above.
Anthropic’s engineering write up describes the same three levels in the same order.
Source: Anthropic, Equipping agents for the real world with Agent Skills, https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
Everything the script argues about which five Skills are worth building first, and about how agents typically fail, is the writer’s opinion and is presented as such in the narration (“I would”, “I think”, “my rule would be”). It is not a measured claim and no source is offered for it.
The following figures appear on screen as illustration of a scenario rather than as measurements of anything, and are deliberately drawn from the invented example repo the video uses throughout rather than from a real project:
079-twelve-files-is-obvious119-retries-forever-happy-path096-apparently-not-obvious051-assumes-the-old-folder044-find-the-commandsNone of these is presented as a statistic about Claude, about any tool, or about developers in general.