OpenAI Quietly Ships Skills in ChatGPT and Codex CLI

OpenAI quietly shipped support for filesystem-based “skills” in both ChatGPT and Codex CLI, aligning with Anthropic’s lightweight approach. In ChatGPT, built-in skills guide robust document/PDF handling by rendering pages to images and iterating for layout fidelity; the author’s test produced a polished PDF after self-correction. Codex CLI now loads user skills from ~/.codex/skills, and successfully used one to scaffold a working Datasette plugin—evidence that skills are practical and worth standardizing.

Key Points

OpenAI has introduced filesystem “skills” in ChatGPT (Code Interpreter) and Codex CLI, mirroring Anthropic’s simple folder-with-Markdown pattern.
ChatGPT’s skills include documents and PDFs, using page rendering to PNGs plus vision models to preserve layout and visuals over raw text extraction.
A real-world test produced a high-quality PDF after iterative self-checks, including automatic font substitution to handle macrons in “kākāpō.”
Codex CLI can load user-installed skills from ~/.codex/skills; Willison used this to have Codex generate a functioning Datasette plugin.
The author calls for a formal, minimal skills spec—potentially overseen by the Agentic AI Foundation—because the pattern is gaining rapid adoption.

Sentiment

The community broadly agrees that skills are a useful and well-designed pattern, with particular enthusiasm for their token efficiency compared to MCP. However, there is notable pushback on novelty claims, with a vocal minority insisting the concept is just rebranded documentation. The tone toward Anthropic is warm and respectful, with many viewing their simple-but-sticky approach as genuinely clever even if not revolutionary.

In Agreement

Skills provide real value through their lazy-loading model: only a one-line index entry per skill is in the base context, and full instructions load only when needed — unlike MCP which forces all tool definitions into every prompt.
The filesystem-based, folder-per-skill pattern is simple enough to implement anywhere and enables cross-platform portability, making it a genuinely useful standardization even if the underlying concept is not new.
Skills are a better abstraction than MCP for many use cases: they work in web UIs, mobile apps, and cloud environments where local MCP servers cannot run.
The pattern of pairing markdown instructions with Python CLI scripts is an effective way to blend LLM reasoning with deterministic, reliable execution.
Anthropic continues to produce deceptively simple innovations (MCP, skills, CLAUDE.md) that gain adoption across the ecosystem, and OpenAI catching up validates the approach.

Opposed

Skills are not a new invention — they are just dynamic prompt extension, a concept that exists in many forms across platforms, and any developer can implement them in 20-30 lines of code.
The framing of skills as a groundbreaking pattern is marketing hype; it is fundamentally just good documentation and context engineering that developers have always done.
Calling for a vendor-neutral skills standard or foundation (like the Agentic AI Foundation) feels premature and performative for something this simple, drawing comparisons to Silicon Valley's satirical 'tethics.'
MCP is already a sufficient packaging mechanism for skills plus code; the distinction between skills and MCP is overstated.