# Will this Update from OpenAI Make AI Agents Work Better?

Source: https://www.youtube.com/watch?v=vjlk3aKZv20
Recap page: https://rapidrecap.app/video/vjlk3aKZv20
Generated: 2025-12-18T01:03:07.998+00:00

---
## Quick Overview

The update from OpenAI, integrating Anthropic's brilliant "skills" mechanism into ChatGPT and Codex CLI, will make AI agents work better by enabling them to dynamically manage files and execute specific procedural knowledge without requiring the user to constantly re-explain context, thereby improving efficiency and reliability over fragmented, custom-built agents.

**Key Points:**
- OpenAI integrated Anthropic's "skills" mechanism into both ChatGPT and Codex CLI, allowing models to dynamically manage files like spreadsheets and PDFs.
- This modular approach creates foundational pieces for next-gen LLM agents, contrasting with older methods that relied on building bespoke agents for every use case.
- Simon Willison noted that this adoption by OpenAI and Google reinforces his belief that Claude Skills are a "bigger deal than MCP," a standard he previously wrote about.
- Skills are structured as folders containing a SKILL.md file with metadata (name, description) and optional code/scripts, enabling progressive disclosure of context.
- The implementation allows for complex, composable resources, moving away from monolithic agents that required hardcoded context or constant re-explanation of processes.
- The structure of Skills (using Markdown and optional scripts) allows for greater efficiency and reliability by loading only necessary context when a skill is triggered.

![Screenshot at 00:04: A screenshot of Simon Willison's tweet announcing that OpenAI is adopting Anthropic's brilliant "skills" mechanism, which is a key indicator of the industry trend toward standardized agent capabilities.](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-00-04.png)

**Context:** The video discusses the recent integration of a "skills" mechanism, originally pioneered by Anthropic for their Claude model, into OpenAI's ecosystem, specifically ChatGPT and the Codex CLI tool. This feature allows Large Language Models (LLMs) to access and execute external instructions, scripts, and resources—packaged in a structured folder format—to perform specific tasks, such as managing files, which was previously a major limitation for building reliable AI agents.

## Detailed Analysis

The main takeaway is that OpenAI has adopted a "skills" mechanism, similar to Anthropic's, allowing models like ChatGPT and Codex CLI to dynamically manage files and execute procedures using structured folders. This addresses the problem where intelligent agents previously lacked specific expertise, were rigid, and required continuous context re-explanation. The skills are implemented as folders containing a SKILL.md Markdown file which defines metadata (name, description) and instructions/scripts. This structure facilitates a "progressive disclosure" model, where Claude only loads the necessary context (the actual body of the skill file) when it deems the skill relevant to the current task, avoiding unnecessary token consumption. Simon Willison views this standardization, evidenced by both OpenAI and Google adopting similar concepts, as a major step forward, suggesting it is a bigger deal than the Model Context Protocol (MCP). This composable approach allows for building reliable, specialized agents without the complexity of building monolithic agents for every single task, fostering greater adoption of AI agents across organizations.

### Comparison to MCP

- Skills offer the same advantage as MCP without needing a new CLI tool
- Skills use a Markdown file to describe the task, adding scripts only if needed for reliability
- MCP, like GitHub's official version, consumes thousands of tokens of context unnecessarily

### The Mechanics of Skills

- A skill is a directory containing a SKILL.md file with YAML frontmatter (name, description) and optional content/scripts
- Metadata is the first level of progressive disclosure; the actual body is the second level
- Claude loads the full skill context only when it deems it relevant to the current task

### Future Implications

- Expect a Cambrian explosion in Skills, making the MCP rush look pedestrian by comparison
- The simplicity and composability allow anyone to build custom agents, not just developers
- This system enables continuous, tangible learning for agents rather than context-dependent, hardcoded expertise

![Screenshot at 00:04: Simon Willison's tweet highlighting that OpenAI is adopting Anthropic's skills mechanism, setting the stage for the discussion on standardized agent capabilities.](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-00-04.png)
![Screenshot at 01:24: A slide from Anthropic's post titled "Equipping agents for the real world with Agent Skills," visually defining what a skill is—an expert in a folder.](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-01-24.png)
![Screenshot at 03:35: A diagram illustrating the concept of bundling additional content \(reference.md, forms.md\) alongside the core SKILL.md file, showing third-level progressive disclosure.](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-03-35.png)
![Screenshot at 03:55: A diagram titled "Skills and the Context Window" showing the sequential operation where Claude decides to "trigger" a skill based on the context window contents.](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-03-55.png)
![Screenshot at 07:34: The main visual graphic from the AI Engineer post contrasting the "Old Way: Monolithic Agents" \(separate agents for each domain, hardcoded expertise\) against the "New Way: Agent + Skills" \(a general agent with many skills, packaged in reusable folders\).](https://ss.rapidrecap.app/screens/vjlk3aKZv20/00-07-34.png)