OpenAI's codex system AGENTS.md file

AGENTS.md

1. When using command line tools whose output can be unpredictably long prefer to pipe it into a temporary file so you can search in it later and save tokens, unless you know command line way to filter output down to just what you need
2. Do not create logs or build artifacts in the project folder unless explicitly asked for, use temporary folders and clean up after yourself unless you need to draw user attention to something that's there, then keep the artifacts

You have two more tools available to you: web search tool and content download tool

Web Search Tool

Overview

The web search tool enables LLMs to retrieve up-to-date information from the internet by querying search engines and returning structured results. It is designed for tasks such as fact-checking, summarizing news, finding recent events, or gathering external knowledge not present in the training data.

Usage

To perform a search, use the following command format:

ddgr --json -n <number_of_results> "<search_query>"

• --json: Outputs results in JSON format for easy parsing.
• -n <number_of_results>: Limits the number of results returned up to 10.
• <search_query>: The search string, enclosed in quotes.

Example

ddgr --json -n 10 "WWDC ios 26 split view"

Response Format

The tool returns a JSON array of search results. Each result contains:

• abstract: A brief summary or snippet from the web page.
• title: The title of the web page.
• url: The direct link to the web page.

Example Output

[
  {
    "abstract": "With iPadOS 26, Apple is introducing major changes to the way that multitasking works...",
    "title": "iPadOS 26 Multitasking Update Gets Rid of Split View and Slide Over",
    "url": "https://www.macrumors.com/2025/06/10/ipados-26-no-split-view-slide-over/"
  },
  ...
]

How LLMs Should Use

• Formulate a clear, concise search query relevant to the information needed.
• Parse the JSON response to extract relevant abstracts, titles, and URLs.
• Use the abstracts for quick answers or summaries.
• Reference URLs for source attribution or deeper research.

Best Practices

• Use specific queries to improve result relevance.
• Limit the number of results to avoid information overload.
• Always cite sources when presenting information from search results.

Content Download Tool

Overview

The content download tool enables LLMs to fetch and extract the main textual content from a given web page URL. It is useful for retrieving full articles, blog posts, or other readable content for summarization, analysis, or citation.

Usage

To download and extract content, use the following command format:

trafilatura -u "<url>"

• -u <url>: Specifies the URL of the web page to extract content from.

Example

trafilatura -u "https://www.macrumors.com/2025/06/10/ipados-26-no-split-view-slide-over/"

Response Format

The tool outputs the main text content of the web page, typically including the article title and body. This content is suitable for summarization, fact extraction, or citation.

Example Output

iPadOS 26 Multitasking Update Gets Rid of Split View and Slide Over
With iPadOS 26, Apple is introducing major changes to the way that multitasking works, and part of that update will see the removal of Split View and Slide Over, two multitasking interface options that have long existed on the iPad.
iPadOS 26 allows for multiple app windows, each of which can be resized freely …

How LLMs Should Use

• Use this tool to retrieve the full readable content of a web page for deeper analysis or summarization.
• Reference the extracted content directly or cite the source URL for attribution.

Best Practices

• Use this tool for pages where full article text is needed, not just summaries.
• Avoid extracting content from pages with little or no readable text.
• Always cite the source URL when presenting extracted content.