How to speed up your coding agent

I've analyzed the leaked system prompts, and here's what I found about how to speed up your LLM.

The prompts are for Claude (Opus 5.5) and OpenAI's Codex desktop app (GPT-6), from the CL4R1T4S collection on GitHub. I was looking for which local tools a coding agent expects to find on the machine.

A lot of checking whether a tool exists, a failed import, a pip install, a retry. Anthropic's cloud sandbox avoids that. Its skills tell the model outright that a library is preinstalled.

You can get the same effect locally. Install these once:

• Search: rg (ripgrep), jq
• Documents: pandoc, LibreOffice, poppler (pdftotext, pdftoppm), qpdf, tesseract
• Python: markitdown, openpyxl, pandas, pypdf, pdfplumber, pymupdf, reportlab, python-pptx, pytesseract, pdf2image, Pillow, lxml, defusedxml
• Node: docx, pptxgenjs, sharp, react-icons, playwright

Then tell the agent what's there. One line in CLAUDE.md or AGENTS.md listing the installed tools removes extra checks at the start of every task. The vendor prompts do exactly this for their own sandboxes.

Two things I did not expect:

1. OpenAI's Codex bundles Node, Python and document libraries and hands the model their paths, and it prefers rg over grep.
2. Claude's prompt tells the model to avoid grep, cat and sed in the shell and use its own tools.

To be precise: the model itself runs at the same speed. The agent finishes sooner because it stops spending turns on setup.