Why Anthropic Switched from Specialized Agents to Skills—and What It Means for AI Development

Background

In the past year the Model‑Connect‑Protocol (MCP) became the de‑facto standard for connecting AI agents to external tools and data sources. Anthropic released Claude Code, a general‑purpose programming agent, together with the Claude Agent SDK that provides a production‑ready, out‑of‑the‑box agent.

Problem

Although agents can reason and execute code, they lack the domain‑specific expertise required for many real‑world tasks. This gap motivated the creation of Agent Skills , a lightweight, file‑based packaging format that bundles workflows, best practices, and scripts so that agents can retrieve and apply specialized knowledge on demand.

What is an Agent Skill?

An Agent Skill is a directory containing a small set of plain‑text files. The minimal structure is:

my_skill/
├── SKILL.md          # full documentation, instructions, and code snippets
├── docs.md           # optional auxiliary documentation
├── slide-decks.md   # optional reference material
└── apply_template.py # optional executable script

The SKILL.md file begins with a YAML header that provides metadata (name, description). At runtime the agent first sees only this metadata; the full content is loaded lazily when the skill is selected, keeping token usage low (metadata ~50 tokens, full skill ~500 tokens, optional references >2000 tokens).

Progressive Disclosure

When the agent decides a skill is needed, it reads the complete SKILL.md. Skills may also contain a references/ sub‑directory with large supporting documents that are fetched only if required.

Skill Types

Basic Skills : Core capabilities such as document, spreadsheet, and presentation handling. Public examples are available at https://github.com/anthropics/skills/tree/main/skills/public.

Partner Skills : Built by third‑party services (e.g., K‑Dense, Browserbase, Notion) to expose their APIs directly to agents.

Enterprise Skills : Proprietary skill sets that capture internal workflows, compliance rules, and institutional knowledge.

Architecture

Agent Loop : The reasoning engine that decides the next action.

Agent Runtime : Execution environment (code interpreter, file system).

MCP Server : Connects the runtime to external tools, databases, and APIs.

Skill Library : Stores the hierarchical skill directories and serves metadata on demand.

This separation allows each layer to evolve independently while keeping the overall system understandable.

Creating a Skill – Example

Suppose you want to add a front‑end design skill that advises on typography, colors, and animation. You would create a directory frontend_design/ with the following files:

frontend_design/
├── SKILL.md          # "name: Front‑End Design", "description: Brand‑compliant UI guidance"
├── apply_template.py # Python script that modifies a PowerPoint file using python‑pptx
└── slide-decks.md   # Markdown that lists color palettes and usage examples

The script can be invoked by the agent as a tool: python apply_template.py my_presentation.pptx When the agent determines that UI advice is needed, it loads SKILL.md, reads the description, and then executes the script only for the relevant presentation file.

Trends

Increasing Complexity : Early skills were simple reference documents; newer skills orchestrate multi‑step workflows, perform heavy calculations, and generate formatted outputs (e.g., a financial model builder that fetches market data, runs Python‑based Excel modeling, and produces a report).

Non‑Developer Adoption : Skill‑creation tools guide product managers and analysts through a wizard‑style workflow, enabling a first skill to be built and tested in ~30 minutes.

Vertical Deployments

Financial Services

DCF Model Builder – constructs discounted cash‑flow models with WACC and sensitivity analysis.

Comparable Company Analysis – generates tables of multiples and benchmarks.

Earnings Analysis – processes quarterly results and creates investment update reports.

Startup Coverage – assembles research reports with integrated financial models.

M&A Due Diligence – runs a standardized merger analysis framework.

Pitch Materials – produces client‑facing presentations following industry standards.

Healthcare & Life Sciences

Bioinformatics Package – wraps scVI‑tools and Nextflow pipelines for genomics and single‑cell RNA‑seq analysis.

Clinical Trial Protocol Generation – accelerates protocol drafting.

Scientific Question Selection – assists researchers in identifying high‑impact questions.

FHIR Development – helps developers write accurate interoperability code.

Prior Authorization Review – cross‑references coverage policies, clinical guidelines, and patient records to reduce administrative burden.

Standardizing Agent Skills

Anthropic is publishing the open specification Agent Skills at https://agentskills.io. The goal is cross‑platform portability: a skill written for Claude should work unchanged with any AI platform that implements the standard.

Key Operational Details

Version Control : Because a skill is just a directory of files, it can be stored in Git, synced via Google Drive, or managed in any existing file‑storage system.

Token Management : Only the YAML header is sent to the model initially. Full skill content is streamed only when the model explicitly requests it, preventing context‑window overflow.

Tool Integration : Scripts inside a skill are executed in the Agent Runtime. They are self‑documenting, modifiable, and do not need to remain in the model’s prompt.

Extensibility : New skills can be added without redeploying the entire agent. The MCP server discovers available skills at start‑up and makes them selectable by the Agent Loop.

Conclusion

Agent Skills turn code and documentation into a universal primitive for extending AI agents. By packaging domain expertise as version‑controlled files, they enable rapid, low‑overhead skill creation by non‑engineers, support progressive disclosure to keep token usage efficient, and foster an ecosystem where skills are portable across platforms. As the community contributes more skills, agents become increasingly capable, reliable, and adaptable to specialized verticals without requiring bespoke model fine‑tuning.