OpenAI Rolls Out Skills Mechanism in ChatGPT and Codex CLI, Enhancing Customization and Functionality

OpenAI has quietly introduced a skills mechanism to both their ChatGPT platform and the Codex Command Line Interface (CLI). This move mirrors similar capabilities seen in Anthropic’s Claude, where skills are implemented as simple folders containing Markdown files and optional resources. The adoption of this feature by OpenAI signifies a significant step towards making large language models (LLMs) more modular and customizable for specific tasks.

Skills in ChatGPT

The new skills mechanism in ChatGPT is particularly interesting because it leverages the Code Interpreter feature to interact with a /home/oai/skills directory. Elias Judin first highlighted this capability on Twitter, noting that you can access these skills by simply prompting:

Create a zip file of /home/oai/skills

When I tried this myself, ChatGPT returned a zip file containing the skills directory. The content includes skills for handling spreadsheets, DOCX files, and PDFs. One notable approach is how OpenAI handles PDFs: they convert each page to a PNG image before processing it with their vision-enabled GPT models. This method helps preserve layout and graphical information that might otherwise be lost during text extraction.

Implementation Details

• Directory Structure: The /home/oai/skills directory contains subdirectories for different skills, each with its own skill.md file.
• PDF Handling: PDFs are converted to PNG images per page before processing, ensuring layout and graphical data are retained.
• Skill Execution: When a user prompts ChatGPT to perform a task using a skill, the model first reads the skill.md file for guidelines, then executes the necessary steps.

For example, I prompted ChatGPT to create a PDF summarizing the current situation of rimu trees and its implications for kakapo breeding season. The process unfolded as follows:

• Reading Guidelines: GPT-5.2 Thinking started by reading the skill.md file for PDF creation guidelines.
• Researching Content: It then searched for information on "rimu mast" and "Kākāpō 2025 breeding status."
• Generating Output: The model took just over eleven minutes to produce the final PDF.

Skills in Codex CLI

The skills mechanism is also available in OpenAI’s Codex CLI, a tool designed for code generation and automation. Similar to ChatGPT, the Codex CLI can access a /home/oai/skills directory containing predefined tasks and guidelines. This feature allows developers to leverage pre-built skills for common coding tasks, enhancing productivity and consistency.

Comparison with Anthropic’s Claude

Anthropic’s implementation of skills in Claude is quite similar, with skills stored as folders containing Markdown files and optional scripts. The key difference lies in the specific tasks and tools each platform supports. For instance, while both platforms handle PDFs by converting them to images, OpenAI’s vision-enabled GPT models offer unique capabilities for processing graphical data.

Implications for Practitioners

The introduction of skills mechanisms in ChatGPT and Codex CLI has several implications for practitioners:

• Customization: Users can create and share custom skills, tailoring the LLMs to specific use cases.
• Modularity: Skills can be easily added or removed, making it simpler to manage and update capabilities.
• Efficiency: Pre-built skills reduce the time and effort required to perform common tasks, enhancing productivity.

Conclusion

OpenAI’s adoption of a skills mechanism in both ChatGPT and Codex CLI marks a significant advancement in the flexibility and utility of large language models. By allowing users to define and share custom skills, OpenAI is empowering developers and researchers to leverage these powerful tools more effectively for a wide range of applications.