---
type: "synthesis"
sources: ["cross-day"]
tags: ["terminology", "architecture", "claude-skills"]
id: "arc-skills-primitive-three-flavors"
---
## The word "Skills" appears in 5 of 6 videos. It does not mean the same thing.

This is the single biggest source of confusion when reading the corpus as a unified body. Below is a decoder.

## Flavor 1 — Anthropic-native Skills (frontmatter + instructions + examples)

The formal Anthropic product feature: a file with a YAML-style frontmatter (name + trigger description), an instructions body, and optional few-shot examples. Lives inside Claude. Travels across every chat where enabled.

- [[concept-claude-skills-d1]] (Alex) — the canonical exposition; emphasizes that the **description matters more than the body** ([[claim-description-importance]], [[contrarian-description-over-instructions]], [[framework-skill-anatomy]]).
- [[concept-claude-skills-d4]] (Sabrina/Kipp, MAG) — same primitive, treated as the unit of *compounding* — note the `update the skill` command that mutates the file ([[framework-skill-refinement-loop]]).

## Flavor 2 — Configured agents installed as JSON / Skill files

A Claude *desktop* feature that lets you pre-load entire SOPs into named agents. Closer to a custom GPT than to a function call.

- [[concept-ai-agent-skills]] (CCC) — the four CCC agents (Creator Finder, Viral Spotter, Transcriber, Knowledge-Base Rewriter) are each installed as JSON Skill files.
- [[concept-agent-skills]] (Sabrina, Day 3) — Skills as *machine-readable documentation* (a `SKILL.md` plus rule files) that teach Claude Code how to use a specific framework. The Remotion skill is installed via `npx skills add remotion-dev/skills`. Invoked **implicitly** by mentioning the framework in natural language ([[quote-implicit-triggering]]).

## Flavor 3 — Project-folder instruction files (informal usage)

A looser, possibly conflated usage: any persistent brand-context document saved in a local folder that Claude Code reads on startup.

- [[concept-claude-code-skills]] (Tim) — the Tim primer itself flags that this *may* conflate (a) user-managed instruction files vs. (b) a model-native named feature. The *pattern* is real; the *product framing* should be verified against current Anthropic docs.

## Why this matters for the downstream agent

When a user asks "how do I make a Claude Skill?" you must ask which flavor:

- For a **frontmatter-based reusable workflow**: route to [[framework-skill-anatomy]] (Alex).
- For an **agent with embedded SOP for a multi-step task**: route to [[concept-ai-agent-skills]] (CCC).
- For **framework-specific developer documentation** Claude Code reads: route to [[concept-agent-skills]] (Sabrina) and [[action-install-remotion-skill]].
- For a **persistent brand-context folder**: route to [[concept-claude-code-skills]] (Tim) — with the caveat that this may be informal terminology.

The Anthropic product surface has clearly grown faster than the vocabulary. Expect this taxonomy to keep drifting.