Skills, Commands, Sub-agents

Summary: Overlapping Abstractions in Anthropic Claude Code and a Proposed Simplification

Clips from the video discuss recent Claude Code changes that have made the product’s three abstractions—slash commands, skills, and sub-agents—increasingly overlap, and proposes a simplification.

What Changed in Recent Releases (notably 2.1.0 and follow-ups)

1) Slash commands and skills are being merged conceptually and in invocation

• Skills now also appear as slash commands, so users can explicitly invoke skills with:
  • /<skill-name>
• Skills still support auto-discovery, but the behavioral differences versus slash commands have been reduced.

2) Skills can now run in their own context window

• A new YAML front matter field/keyword is mentioned (subtitles suggest context_fork).
• This lets a skill execute as if it were a sub-agent, using a separate context window so it doesn’t consume the main context.

3) Skills can list usable “agents”

• Another YAML front matter field (agent) allows a skill to specify additional built-in agents it can use (e.g., explore, plan, etc.).
• The speaker notes this part is confusing because the semantics aren’t clear.

Current Meaning of the Three Abstractions (as explained)

1) Slash Commands

• Treated as saved prompt snippets
• User explicitly triggers them by typing /command-name
• Purpose: avoid retyping recurring prompts

2) Skills

• An evolution of slash commands: still Markdown, but packaged as a folder
• A skill folder typically includes:
  • a skill.md (Markdown prompt)
  • additional resources (e.g., code snippets and HTML)
  • examples include content like PDF manipulation or generating Word documents
• Auto-discovery mechanism:
  • Claude inspects YAML front matter/metadata first
  • Uses progressive disclosure:
    • only the front matter is visible initially, even if the full skill.md is long
  • The full skill content is loaded only if needed, reducing main-context usage

3) Sub-Agents

Two historical value propositions:

1. Persona/role definition
   • e.g., “expert code reviewer” or “security engineer” with instructions
2. Execution in a separate context window
   • historically used to prevent sub-agents from bloating the main context

Previously, the separate-context capability was mainly why sub-agents existed.

Why the Model Is “Muddied” Now

After the changes:

• You can invoke skills like slash commands → reduces the distinction.
• You can run skills in their own context → reduces the unique advantage of sub-agents.

Result: the system’s abstractions increasingly overlap and are less mutually exclusive.

Proposed Simplification: One Unifying Abstraction (Skills)

The speaker argues the current three abstractions mix two orthogonal concerns:

1. Type of knowledge being encoded

   • Conceptual/domain knowledge
     • historically sub-agent oriented (persona-like guidance)
   • Procedural knowledge
     • historically skill oriented (step-by-step behaviors and/or code/resources)

2. Where it executes

   • Main context window vs separate context window
   • historically sub-agent oriented for the separate-context part

Proposal

• Collapse everything into one primitive: Skills
• Keep a single switch (e.g., front matter such as context_fork) to choose:
  • run in the main context window or separate context window
• Describe personas inside the skill content (e.g., in skill.md) rather than requiring a separate “sub-agent” abstraction
• Handle composition via references:
  • Skills can explicitly specify other skills/agents to consider
  • Auto-discovery can still pull in relevant skills from the skills directory

“Guide / Reference” Mention

The speaker references a quick reference / decision tree:

• If the goal is to capture autodiscovered knowledge/capabilities, choose skills
• If you need a persona, encode it as a sub-agent

They emphasize these distinctions have mostly disappeared due to the merging changes.

Main Speakers / Sources

• Main speaker: The YouTube narrator (not named in the subtitles)
• Source referenced: Claude Code / Anthropic release updates and change logs (the speaker discusses versions like 2.1.0 and subsequent changes)