Context efficiency
What stays in the prompt?
This page compares two ways to use skills in Codex: installed Codex skills on their own, and skills delivered to Codex through @skills.
Codex without @skills
Installed skills contribute lightweight metadata up front. Full instructions load after Codex selects a skill.
Codex with @skills
@skills delivers a skill to Codex when its address appears in the prompt. Unreferenced skills add no resident context.
The tradeoff: native Codex skills make automatic discovery convenient; adding @skills gives Codex a lower-context, explicitly addressed path.
| Codex + @skills mode | What happens | Resident context |
|---|---|---|
| Reference | Call a path such as @skills:gh:owner/repo. | None until use |
| Saved | Save locally for faster reuse and autocomplete. | None until use |
| Installed | Add metadata for automatic routing. | Metadata only |
Simulator 01
Context cost inside Codex
Compare installed Codex skills with skills referenced through @skills.
Illustrative model: 40 metadata tokens per installed Codex skill and a 1,500-token execution payload. Referenced @skills content is counted only when used.
Workflow reliability
How does the agent choose each step?
Without explicit addresses, Codex must select the right installed skill at every step. With @skills, the prompt can name each dependency directly.
Codex without @skills
Codex matches each task to installed skill metadata. It is flexible, but every added routing decision introduces another chance to misroute.
Codex with @skills
The prompt names the skill addresses for Codex in advance. Selection is deterministic, though execution can still fail for other reasons.
Simple mental model: @skills gives Codex file-import-style addressing. If you know the dependency, name it directly.
Simulator 02
Compounding routing risk
Adjust workflow length and per-step routing reliability.
This isolates skill-selection risk. The 100% line means @skills selected the intended address for Codex—not that the overall task is guaranteed to succeed.