The Open-Standard Era: What Skill Authors Need to Know in Mid-2026
If you author skills, the ground under your work changed , quietly and almost entirely in your favor, over the past year. The format you write is no longer a single vendor’s feature. It is a published open standard with real cross-vendor adoption, and that changes what "good authoring" means. Here is the short version, with dates, and what to do differently.
The timeline that matters
- October 2025: Anthropic introduced Agent Skills (the SKILL.md format, frontmatter that signals when a skill applies, and progressive-disclosure loading that keeps unused skills cheap).
- December 2025: the format was published as an open standard, with the specification and SDK at agentskills.io, free for any agent product to adopt.
- Since then, adoption has spread well beyond Anthropic: Microsoft, OpenAI, GitHub, Cursor, Figma, and Atlassian are among the adopters, and the same SKILL.md files are read by tools including OpenAI Codex, GitHub Copilot, Gemini CLI, and VS Code. Official skill collections now ship under more than one company’s name.
- Alongside it, Claude Code folded custom slash commands into the skills system (one mechanism instead of two), and the Model Context Protocol, the connective layer skills lean on to reach external tools, moved to foundation governance under the Linux Foundation.
None of those is a flashy capability demo. They are decisions about who owns the format and how portable your work is, which, if you are investing real hours in writing skills, matters more than any single feature.
What it changes for authors
When a skill ran in one tool, sloppy habits were survivable: a hard-coded path worked because there was only one machine, an unstated prerequisite was fine because there was only one environment. In the open-standard era your skill will be read by agents you did not plan for, on machines you have never seen. Five habits cover most of the gap:
- Reference bundled files by relative path, always. An absolute path is a promise about someone else’s filesystem.
- Name every prerequisite in the body: tools, credentials, services. The reader may be an agent with no way to guess.
- Keep bundled scripts dependency-free. The standard travels; your pip installs do not.
- Treat the description as a public API. Once others install your skill, broadening the trigger reroutes THEIR requests. Version it and note behavior changes.
- Test in more than one tool if you distribute. The file format is guaranteed portable; instruction-following fidelity varies by model. A small trigger table re-run per tool is cheap insurance.
The flip side: read what you run
The same openness that makes your skills portable makes everyone else’s skills installable, and a skill is instructions your agent will follow, sometimes with scripts it will execute. The ecosystem has no gatekeeper, which means the reading discipline is yours. Before any third-party skill lands in your directory, read the whole SKILL.md and every bundled script. Our five-minute audit walks through exactly that.