Too many skills? Codex trims their descriptions to fit 2% of context, so put the trigger words first

#What changed

Expo published eval results, built with Georgian AI Lab, on how coding agents find its skills. In Codex, the starting list of skills (each skill's name, description, and file path) shares one budget: at most 2% of the model's context window, or 8,000 characters when the window size is unknown. OpenAI's Codex skills docs confirm this cap. When the list overflows, Codex shortens descriptions first and, with very large skill sets, can leave some skills out of the list with a warning.

In one Expo test setup with 175 skills installed (24 of them Expo's) and a 272k window, every Expo skill was still listed, but only about the first 60 characters of each description survived.

Expo also found that adding a single entrypoint skill, expo-overview, which routes to its more specialized skills, raised the share of Claude Code sessions loading a relevant Expo skill from about 9% to 55% across 10 app-building tasks. Recall of all the relevant skills rose only from 2% to 18%.

#Who is affected

Anyone using Codex CLI, the Codex IDE extension, or the ChatGPT desktop app with many skills or plugins installed, and anyone writing skills or plugins for other people.

#What to do now

• Front-load every skill description: put the use case and trigger words in roughly the first 60 characters, and give the skill a descriptive name.
• Turn off skills you don't use with a [[skills.config]] entry set to enabled = false in ~/.codex/config.toml, then restart Codex, so they stop crowding the list.
• If you ship a bundle of related skills, add one entrypoint skill whose description says to start there and whose body maps common goals to the other skills.
• For skills that must run, invoke them by name with $skill-name instead of relying on implicit matching.
• Test discovery with realistic prompts that don't name the skill, in a catalog as crowded as your real setup.

#What is not confirmed

The 9% to 55% and 2% to 18% figures are Expo's own measurements on Claude Code (Sonnet), 10 tasks with three runs each. They have not been independently replicated and may not carry over to Codex. The 60-character figure comes from one Expo test environment; the actual cut depends on your context window and how many skills you have. The 2% cap itself is documented by OpenAI.

Sources: Expo on X, Expo blog: Your skills need an evaluation mechanism, OpenAI Codex skills docs.