The Agent Stdlib Is a Folder of Markdown

Anthropic's skills repository, anthropics/skills, has surpassed 170,000 stars and 20,000 forks on GitHub, becoming the de facto standard for packaging agent expertise through SKILL.md files. The format, introduced by Anthropic in October 2025 and released as an open standard in December 2025, has been adopted by major AI clients including GitHub Copilot, OpenAI's Codex, and Gemini CLI, marking a rare cross-vendor consensus. The repo's design, based on progressive disclosure, minimizes context usage and has outpaced alternatives like system prompts and MCP for agent capability management.

AI https://sourcefeed.dev/c/ai Article The Agent Stdlib Is a Folder of Markdown Why anthropics/skills keeps trending: SKILL.md became the default way to package agent expertise in under a year. Mariana Souza https://sourcefeed.dev/u/mariana souza anthropics/skills https://github.com/anthropics/skills is back on GitHub's trending page, now past 170,000 stars and 20,000 forks. That's a strange trajectory for what is, mechanically, a folder of Markdown files. But the repo keeps trending because it turned into something nobody quite planned: the reference library for the format that won the agent-packaging war in under a year. Quick recap for anyone who skipped late 2025. Anthropic shipped Agent Skills in October 2025 as a Claude feature: a skill is a directory with a SKILL.md file YAML frontmatter with a name and description , then freeform instructions , plus optional scripts/ , references/ , and assets/ folders. Simon Willison called it "maybe a bigger deal than MCP" the day it launched, which read as provocation at the time. Then in December Anthropic released the format as an open standard at agentskills.io https://agentskills.io , and the provocation became consensus. The spec's client showcase now lists GitHub Copilot, VS Code, Cursor, OpenAI's Codex, Gemini CLI, Goose, OpenCode, JetBrains' Junie, Databricks, Snowflake, and dozens of smaller agents. OpenAI and Google adopting an Anthropic-invented format is not a common event. It's happened exactly twice, and the other time was MCP. Why a folder of Markdown beat everything else The design bet behind skills is that agent capability is mostly a context-management problem. The spec's core mechanism, progressive disclosure, works in three stages: at startup the agent loads only each skill's name and description a few dozen tokens , it reads the full SKILL.md when a task matches, and it executes bundled scripts or loads reference files only as the work demands. A library of 50 skills costs you almost nothing until one fires. Compare that to the alternatives developers were actually using in mid-2025. Stuffing procedures into the system prompt or a CLAUDE.md burns context on every request whether relevant or not. Fine-tuning bakes knowledge in but needs a training run per revision. RAG needs retrieval infrastructure and still hands the model loose paragraphs instead of a procedure. MCP https://modelcontextprotocol.io solves a different problem entirely: it connects agents to external systems, and its tool schemas sit resident in context, which gets expensive fast. Skills occupy the gap all of those left open. They're version-controlled, diffable in code review, portable across vendors, and free until used. That's also why the standard spread so fast. There was no protocol to implement, no server to run. A client "supports skills" the moment it scans a directory and injects frontmatter into its system prompt. Most agent vendors could ship support in a week, so they did. What the repo is actually for Four parts of the repo carry the weight. The skills/ directory holds Anthropic's examples across design, development, and enterprise workflows. The template/ directory is the fastest honest starting point for writing your own. A spec/ folder tracks the open specification. And the document skills docx , pdf , pptx , xlsx are the production code paths that power Claude's own file generation, published source-available while the rest of the repo is Apache 2.0. That license split matters if you plan to fork: the example skills are yours to remix, the document skills are yours to read. Adoption in Claude Code is two commands: /plugin marketplace add anthropics/skills /plugin install document-skills@anthropic-agent-skills Or skip the plugin system and drop folders into ~/.claude/skills personal or .claude/skills per-repo, shared with your team through git . The same folder works in Copilot, Cursor, and Codex, each with its own discovery path documented in their respective docs https://agentskills.io/clients . The bigger value is pedagogical. Before this repo, "write good instructions for an agent" was folklore. Now there's a browsable corpus showing the pattern that separates skills that work from prompt dumps that don't: a short SKILL.md that reads like a decision procedure, deterministic work pushed into scripts the model runs instead of improvising, and heavyweight reference material split into files loaded on demand. The single highest-value habit to steal is how the mature skills write their description field. It's the only thing the model sees at startup, so it has to function as a routing condition "use when the user asks X or Y" , not a marketing blurb. The part nobody has solved Here's the caution that belongs in every skills article and appears in few. A skill is an unsandboxed prompt plus code your agent will execute with your permissions. Installing one from a stranger's marketplace is curl | bash with extra steps, except the payload can also steer the model itself, which makes skills a prompt-injection vector on top of a code-execution one. There's no signing, no lockfile, no dependency resolution. The community marketplaces springing up around the format are recreating npm's trust problem without npm's decade of tooling. Anthropic's own README is candid that the examples are demonstration-grade and need testing before production use. Skills also don't compose formally. Two skills with overlapping trigger descriptions can fight over the same request, and your debugging tool is reading transcripts. Anyone who maintained a 30-skill library through a model upgrade knows the instructions tuned for one model's behavior can quietly misfire on the next. None of that changes the verdict. This is a genuine shift, not hype: the simplest plausible design won because the constraint that binds agents is context, not connectivity. What developers should do about it is equally simple. Keep your team's operational knowledge as skills in your repos, steal structure from anthropics/skills rather than inventing your own conventions, and read every third-party skill before you install it, the way you claim you read shell scripts. The format standardized in a year. The supply chain around it hasn't started. Sources & further reading - anthropics/skills https://github.com/anthropics/skills — github.com - Agent Skills Overview https://agentskills.io — agentskills.io - Claude Skills are awesome, maybe a bigger deal than MCP https://simonwillison.net/2025/Oct/16/claude-skills/ — simonwillison.net - Agent Skills https://simonwillison.net/2025/Dec/19/agent-skills/ — simonwillison.net - Agent Skills: Anthropic's Next Bid to Define AI Standards https://thenewstack.io/agent-skills-anthropics-next-bid-to-define-ai-standards/ — thenewstack.io Mariana Souza https://sourcefeed.dev/u/mariana souza · Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. Discussion 0 No comments yet Be the first to weigh in.