ClaudeAgents

Building agents with Skills: Equipping agents for specialized work

January 22, 2026 at 12:00 AM

Summary

TL;DR

Anthropic is shifting from building many specialized agents to building “Skills”: versionable file bundles that encode domain workflows, docs, and scripts. Skills load via progressive disclosure so large libraries stay usable within a model’s context window.

What actually happened

Anthropic identified a gap: agents can execute, but often lack domain expertise for real work.
They introduced Agent Skills as organized collections of files (workflows, best practices, scripts).
Skills are designed to be simple: files you can store/share and version with existing tooling.
They describe an architecture layering agent loop, runtime, MCP servers, and a skills library.
They’re publishing Agent Skills as an open standard intended to be portable across platforms.

Key numbers

Metadata shown initially uses ~50 tokens.

Full SKILL.md content is ~500 tokens when loaded.

Reference files can be 2,000+ tokens, loaded only on demand.

“Status report writer” skill: ~100 lines.

“Financial model builder” skill: ~800 lines.

“RNA sequencing pipeline” skill: 2,500+ lines.

Why this was hard

General-purpose agent capability didn’t translate into domain-specific reliability.
Organizations need a way to inject internal context and workflows into agents.
Context windows constrain how much procedural knowledge you can preload.
Traditional “tools” can have bad instructions and inflate prompt context.

How they solved it

Represent expertise as a skill folder: SKILL.md plus optional docs and scripts.
Use YAML frontmatter for skill metadata (name/description) surfaced at runtime.
Apply progressive disclosure: metadata first, SKILL.md if selected, references/ only when needed.
Treat scripts inside a skill as tools (e.g., a Python script to apply slide styling).
Combine skills with MCP servers so a skill can orchestrate external systems and data sources.
Encourage non-engineers to author skills; a “skill-creator tool” supports first skills under 30 minutes.

Why this matters beyond this company

“Files as the primitive” makes domain knowledge versionable, reviewable, and shareable like code.
Progressive disclosure is a practical pattern for scaling knowledge bases under context constraints.
Packaging procedures as scripts can reduce repeated ad-hoc code generation inside agent runs.

Stealable ideas

Put domain workflow + guardrails in SKILL.md, and keep deep docs in references/.
Use frontmatter metadata so agents can choose skills without loading full text.
Promote repeated one-off scripts into maintained skill tools.
Treat agent architecture as layers: loop, runtime, connectors (MCP), and skills.

Read original article

Building agents with Skills: Equipping agents for specialized work

Summary

What actually happened

Key numbers

Why this was hard

How they solved it

Why this matters beyond this company

Stealable ideas

More from Claude

Building agents with Skills: Equipping agents for specialized work

Summary

What actually happened

Key numbers

Why this was hard

How they solved it

Why this matters beyond this company

Stealable ideas

More from Claude