---
type: "synthesis"
spans_days: [1, 2, 3, 4, 5]
tags: ["skills", "terminology", "architecture"]
id: "arc-skills-semantic-drift"
sources: ["cross-day"]
---
## What this arc tracks

Five of the six videos in the series invoke a feature called "Skills," but every speaker means something subtly different. A downstream agent must hold all five definitions simultaneously to avoid talking past users.

## The five definitions

- **Day 1 — Alex (Grow with Alex):** A [[concept-claude-skills-d1]] is a **portable text file** with three layers — frontmatter (trigger description), instructions block, optional examples. See [[framework-skill-anatomy]]. The defining feature is that the *description* is the routing key. See [[claim-description-importance]] and [[quote-skill-definition]].
- **Day 2 — Alessio (CCC):** An [[concept-ai-agent-skills|AI Agent Skill]] is a **JSON file with a strict Standard Operating Procedure** installed into Claude desktop. Each Skill is narrowly scoped (Creator Finder, Viral Spotter, etc.) and chains with others in a pipeline ([[framework-ccc-content-pipeline]]).
- **Day 3 — Sabrina (Claude Code):** An [[concept-agent-skills|Agent Skill]] is a **directory of machine-readable documentation** (a `SKILL.md` plus rule files) installed via `npx skills add remotion-dev/skills`. Triggered *implicitly* by mentioning the framework. See [[quote-implicit-triggering]] and [[action-install-remotion-skill]].
- **Day 4 — Sabrina (Co-Work):** A [[concept-claude-skills-d4|Claude Skill]] is a **reusable instruction pack invoked by slash command** (`/write-content`), built via [[concept-brand-voice-interview|reverse-engineered interview]], and crucially *mutable* via the command "update the skill with everything we've talked about." Mutability is the source of compounding.
- **Day 5 — Tim (Speaker 1):** [[concept-claude-code-skills|Claude Code Persistent Skills]] are a **local folder of brand context and operational instructions** that the AI reads on each invocation. The speaker may be conflating user-managed instruction files with a model-native feature (see the validation caveat in that note).

## What converges

All five definitions share three structural commitments:
1. **Named, addressable, reusable.** Not a prompt you type again — a thing you invoke.
2. **Context-pre-loaded.** Brand voice / procedure / SOP is captured once.
3. **Portable across sessions.** Lives outside the chat history.

## What diverges

| Axis | Day 1 | Day 2 | Day 3 | Day 4 | Day 5 |
|---|---|---|---|---|---|
| Form | text file | JSON | folder + MD | instruction pack | local folder |
| Triggering | description match | explicit chain | implicit mention | slash command | name reference |
| Mutability | rewrite file | rewrite JSON | edit MD | `/update the skill` | rewrite folder |
| Granularity | one task | one agent | one framework | one voice | one workspace |

## Why this matters

When a downstream user asks "how do I build a Claude Skill?" the answer depends on **which product surface they're using**. Day 1's anatomy applies to the web/desktop app; Day 3's `npx skills` syntax applies to the CLI; Day 4's slash-command pattern applies to Co-Work. They are *not* interchangeable. The most durable abstraction is Day 4's compounding-asset framing — see [[arc-skill-mutability-compounding]] — because it survives any specific product UI.