- Published on
I Tested SKILL.md on 4 of Them, and Cursor Collapsed
- Authors
- Name
- DeepHorus
AI coding agents have promised a universal standard where a single instruction file works across all platforms, but a recent test reveals significant inconsistencies. An independent experiment tested a 40-line SKILL.md file on four popular AI coding agents without modifications. Claude Code, OpenAI Codex, and Gemini CLI executed the task cleanly, while Cursor failed, requiring extensive debugging. This highlights the gap between industry promises and real-world interoperability.
SKILL.md is an open file format introduced by Anthropic on December 18, 2025, under the Apache 2.0 and CC-BY-4.0 licenses. The specification, published at agentskills.io, aims to create a common language for AI coding agents. It includes a reference SDK and validation tooling to ensure consistency across implementations.
The adoption of SKILL.md has been remarkably rapid. In under 90 days, the number of tools supporting the format grew from one registry to between 26 and 32. Vercel's skills.sh leaderboard now tracks over 90,000 published skills across 19 agents, while SkillsMP indexes more than 1.6 million public SKILL.md files. This pace makes it one of the fastest standard adoptions in developer tools history.
The implications of these inconsistencies are significant. If agents interpret the same SKILL.md file differently, developers cannot trust that skills will work across tools. This undermines the core promise of 'write once, run anywhere.' The failure of a major agent like Cursor suggests that the standard may require stricter enforcement or better compliance testing.
Reactions from the developer community have been mixed. Some praise the rapid adoption and the open nature of the standard, while others express concern about fragmentation. The experiment's author noted that the gap observed should be more worrying than any benchmark, as it directly impacts developer productivity.
Looking ahead, the success of SKILL.md will depend on how quickly the ecosystem converges on a consistent implementation. Anthropic and other stakeholders may need to invest in certification programs or automated testing to ensure all agents handle the format reliably. Without this, the standard risks becoming another fractured tool in the developer's kit.
