Why Anthropic Introduced Agent Skills and How They Could Transform AI Agents

Motivation

Claude Code has seen rapid adoption for code generation and building complex agent applications. Anthropic, the originator of the Model‑Centered Prompt (MCP) standard, announced a new concept called Skills and plans to release an open standard for Agent Skills in December 2025.

Limitations of MCP

Context‑window inflation and heaviness Developers report that MCP often consumes excessive context‑window space. For example, attaching a Playwright MCP requires injecting the entire Playwright API definition, parameter structures, and even the current DOM tree into the prompt. In a typical MCP workflow, the model acts as a data router, causing intermediate results to consume additional tokens. An official example asks an agent to “download my meeting notes from Google Drive and attach them to a Salesforce lead.” The MCP flow reads the full 20 000‑word transcript into the model’s context, then writes it back to Salesforce, doubling token usage, risking context overflow, and reducing robustness due to hallucinations or truncation.

Token consumption doubles (e.g., a 2‑hour sales meeting transcript may contain 50 000 tokens; the round‑trip adds another 50 000+ tokens).

Context overflow can abort the workflow.

Robustness declines because the model manually “copies‑pastes” large data, increasing hallucinations and formatting errors.

Missing positioning: split between tools and methods MCP standardizes how tools are invoked (the “hand”) but does not define how business logic is orchestrated (the “brain”). Platforms such as Dify or n8n can integrate MCP as a tool provider, yet the actual SOP remains scattered in the model’s implicit knowledge or hard‑coded workflows.

Skill Concept

A Skill represents the ability to complete a specific task, analogous to a chef knowing which utensil to use at each step. Skills are standardized SOPs that capture task definitions, execution flow, and supporting resources.

Previously, developers hard‑coded business flows using Python/JavaScript or drag‑and‑drop YAML (e.g., Dify).

With improved model understanding, these definitions can be expressed in natural language, allowing non‑programmers to codify expertise. Skills engineer and standardize this capability, making it an atomic, callable unit for agents.

Design Philosophy: Progressive Disclosure

Uncontrolled injection of heterogeneous information into a model’s context harms performance. Skills adopt a “professional division, information isolation, on‑demand usage” strategy, presenting only the most necessary information when required.

The Skill architecture consists of three layers:

Metadata layer (always loaded): name and brief description, consuming minimal tokens.

Instruction layer ( SKILL.md, loaded on trigger): core workflow and steps, moderate token cost.

Resource layer (loaded on demand): scripts, reference documents, templates, which may be token‑heavy.

Standard Skill Structure

my-skill/
├── SKILL.md          # [required] core brain with metadata and natural‑language instructions
├── scripts/          # [optional] executable code (Python, Bash, etc.)
├── references/       # [optional] knowledge base (API docs, manuals)
└── assets/           # [optional] static resources (templates, images, sample data)

Benefits

Token economy and data isolation: the model first reads only metadata, then loads the instruction layer, and finally accesses heavy resources only when needed, preventing context bloat and solving the token‑waste issue demonstrated in the Google Drive example.

Robustness improvement: encapsulating complex logic in scripts reduces hallucinations and errors, similar to how tools like Manus or Cursor separate high‑frequency operations from model reasoning.

Using Skills

Running a Skill is decoupled. The standard method uses the Claude Code CLI; environments such as Cursor or VS Code (with Codex) can follow the same file protocol.

Place the skills folder under .claude/skills in the user or project root to activate the Skill.

An example Skill that “cooks braised pork” shows that development and management remain straightforward.

Early Implementation

An early, enterprise‑grade implementation called skill0.atypica.ai demonstrates the Skill concept in practice, providing a platform where agents can manage abilities, make judgments, and move toward engineering‑level reuse.

Specification

Official Skill specification: https://agentskills.io/specification