When Anthropic launched Skills in October, the announcement read like a niche developer feature. Two months later, OpenAI has adopted the same architecture—and the quiet convergence reveals something significant about where AI agents are heading.
Skills are deceptively simple: folders containing Markdown files that tell AI systems how to perform specific tasks. But their adoption by both major AI labs suggests the industry has found a common answer to a fundamental question: how do you make AI assistants consistently good at specialized work?
What OpenAI Just Did
Developer Elias Judin discovered OpenAI’s implementation on December 12 while experimenting with ChatGPT’s Code Interpreter. By prompting the model to create a zip file of its /home/oai/skills directory, he found folders for PDFs, spreadsheets, and documents—each containing instruction files structurally identical to Anthropic’s specification.
The same architecture appeared in OpenAI’s Codex CLI tool two weeks earlier, through a pull request titled “feat: experimental support for skills.md.” The implementation mirrors Anthropic’s approach: skills live in a local directory (~/.codex/skills), each defined by a SKILL.md file with metadata and instructions.
OpenAI hasn’t formally announced the feature. But its presence in both ChatGPT and Codex suggests a deliberate strategy rather than an experiment.
Why Skills Matter
The traditional approach to making AI better at specific tasks involved fine-tuning—expensive, time-consuming model training on specialized data. Skills offer a lighter alternative: instructions and resources that load only when relevant.
Anthropic’s engineering team described the design principle as “progressive disclosure.” Each skill takes only a few dozen tokens when summarized, with full details loading only when the task requires them. This solves a practical problem: context windows are valuable real estate, and stuffing every possible instruction into every request wastes resources.
The architecture works because modern AI models can read and follow instructions dynamically. A skill for PDF handling might include preferred libraries, edge case handling, and output formatting—information the model needs only when processing PDFs.
The Convergence Story
OpenAI adopting Anthropic’s approach isn’t unusual in isolation. AI labs regularly learn from each other’s published work. What’s notable is the structural identity: same file naming conventions, same metadata format, same directory organization.
This compatibility could mean skills written for Claude Code can work with OpenAI’s Codex CLI, and vice versa. Developers could share skills on GitHub like npm packages. The ecosystem becomes interoperable rather than fragmented.
The timing coincides with broader standardization efforts. Anthropic donated the Model Context Protocol to the Linux Foundation on December 9, and both companies co-founded the Agentic AI Foundation alongside Block. Google, Microsoft, and AWS joined as members.
The foundation will steward MCP, Block’s goose project, and OpenAI’s AGENTS.md specification. Skills fit naturally into this standardization push—reusable capability modules that work across platforms.
What This Means for AI Coding Tools
The skills architecture matters most for AI coding tools, where specialized knowledge dramatically improves output quality. A skill for React development might specify component patterns, state management preferences, and testing conventions. A skill for database migrations could include safety checks and rollback procedures.
AI coding startups like Cursor have built businesses on making AI more useful for specific development tasks. The skills framework gives model providers a standardized way to offer similar customization—potentially threatening or complementing third-party tools depending on execution.
For enterprise developers, interoperable skills mean institutional knowledge becomes portable. A company’s internal coding standards, security requirements, and workflow preferences can be encoded once and applied across whatever AI tools the team uses.
The Strategic Subtext
OpenAI’s adoption carries strategic implications. The company has historically favored proprietary approaches—GPT Actions, custom GPTs, platform-specific integrations. Skills represent a pivot toward open standards that work across tools.
One interpretation: OpenAI recognizes that developer ecosystems matter more than proprietary lock-in at this stage. If skills become standard, being compatible matters more than controlling the specification.
Another interpretation: competing with Anthropic’s developer experience requires matching its features. Claude Code has grown aggressively, reaching $1 billion in annualized revenue and integrating into Slack. Skills are part of what makes Claude Code useful; OpenAI needed to respond.
The truth likely includes both factors. AI labs compete intensely on benchmarks and capabilities while collaborating on infrastructure standards that benefit everyone. Skills fall into the second category.
What Comes Next
The immediate opportunity is a skills marketplace—GitHub repositories where developers share specialized instruction sets for common tasks. Anthropic already has an anthropics/skills repository. Expect OpenAI to follow, and expect community-contributed skills to proliferate.
The longer-term question is how deeply skills integrate into AI products. Currently, they’re mostly relevant for developers using CLI tools. But the same architecture could power customization in consumer products—personalized writing assistants, specialized research tools, domain-specific chatbots.
For now, the convergence on skills represents something rare in AI: competing companies agreeing that standardization serves everyone. Whether that cooperation extends to other contentious areas—safety standards, capability disclosures, deployment guidelines—remains uncertain.
But for developers building on AI platforms, the message is clear: skills are becoming infrastructure. Learning to write them now means being prepared for how AI tools will work tomorrow.
Source: https://www.unite.ai/