Anthropic Releases Summary of Experience in Building Claude Code Skills
On June 3, Anthropic published a blog titled "Lessons from building Claude Code: How we use skills," clarifying that the core of Skills is not a single Markdown file, but a complete folder structure.
This folder includes SKILL.md instructions, scripts/ scripts, references/ reference materials, assets/ template resources, config.json configuration, as well as Hooks, logs, SQLite, and other persistent data. The Agent achieves progressive context disclosure through automatic discovery and exploration of the folder, significantly reducing hallucinations and saving tokens.
This framework has been widely applied internally at Anthropic, driving the evolution of Claude Code from simple chat to structured, reusable Agent workflows, allowing developers to directly replicate the folder model to build custom capabilities.
Source: Public Information
ABAB AI Insight
Anthropic has been continuously iterating on the Claude Code plugin and skills system, having previously open-sourced the skills repository and provided templates. This blog systematically summarizes hundreds of internal Skills usage experiences, continuing its path from basic prompt engineering to engineered Agent infrastructure, similar to the early expansion of Claude Artifacts functionality.
On the capital front, Anthropic accelerates developer ecosystem adoption and collects feedback through the free open-source Skills specifications and example folders, while binding Claude Code user stickiness and subscription conversion, transforming internal toolchain advantages into platform-level competitive barriers, paving the way for the commercialization of multimodal Agents.
The evolution of skills/plugin systems in tools like Cursor Composer and Replit Agent, as well as the early IDE extension market; currently, Claude Code is at a mature stage of transitioning Agent toolchains from experimentation to standardized, composable workflow control.
Essentially, this represents a technological substitution and industrial chain reconstruction: the folder-based Skills structure replaces scattered prompts and context management, reshaping the AI development productivity chain through Agent automatic exploration and persistence mechanisms, shifting pricing power from single model calls to platform players mastering reusable skill assets, persistent states, and progressive disclosure frameworks.
ABAB News · Cognitive Laws
A single file is a trap; a folder is a weapon: the more structured assets, the fewer Agent hallucinations.
Context is not a consumption but a compounding: progressive disclosure allows each call to stand on the shoulders of history.
Tools define the boundaries of capability: whoever standardizes the skill folder first will lock in the next generation of the Agent ecosystem.