Codex

Install Codex skills, choose the right scope, invoke workflows reliably, and keep descriptions discoverable as your catalog grows.

Install one skill or the full library

Install one workflow globally:

$npx skills add tushaarmehtaa/tushar-skills --skill remove-ai-slop -g -a codex -y

Or install every catalog skill globally:

$npx skills add tushaarmehtaa/tushar-skills -g -a codex -y

Choose global or project scope

Global skills are available across projects from ~/.agents/skills. Project skills live in .agents/skills and can be reviewed with the repository.

These are the runtime-facing destinations observed from the generated npx skills commands: skills CLI 1.5.23 installed the project and global package into the shared .agents/skills location on August 28, 2026.

Omit -g for project scope:

$npx skills add tushaarmehtaa/tushar-skills --skill remove-ai-slop -a codex -y

Invoke the workflow

Explicit invocation: $remove-ai-slop or /skills

Natural-language requests can also trigger a skill when its description matches. Use explicit invocation when you want deterministic selection.

Keep descriptions discoverable under catalog pressure

Codex initially exposes each skill's name, description, and path. OpenAI caps that catalog at 2% of the model context window, or 8,000 characters when the context size is unknown. Codex shortens descriptions first and may eventually omit skills.

In our September 2–3, 2026 controlled run with Codex CLI 0.152.1 and gpt-5.6-sol, the final capture started with 109 visible skills. Adding up to 100 project probes kept all 209 names visible, but progressively shortened their descriptions:

Project skills added	Total visible	Visible description length
10	119	118–120 characters
25	134	98–100 characters
50	159	72–76 characters
100	209	40–44 characters

Put the distinctive capability and strongest trigger in the opening clause. Do not rely on qualifiers near the end. In a small exploratory check at the 100-probe condition, an opaque front-loaded trigger selected the intended skill in 3/3 runs; the same trigger at the shortened-away tail succeeded in only 1/3 and twice selected the wrong skill. Explicit invocation succeeded in 3/3.

If Codex warns that descriptions were shortened, disable unused skills or plugins and invoke important workflows explicitly. The full SKILL.md still loads after selection. See OpenAI's skill documentation for the catalog policy.

Update and remove

Update the global copy:

$npx skills update -g remove-ai-slop

Remove it only from Codex:

$npx skills remove remove-ai-slop -g -a codex -y

Reload after changes

Codex detects skill changes and newly installed skills automatically. Restart Codex only if an update does not appear.

If discovery still looks stale, verify the folder contains SKILL.md at its top level before restarting the runtime.

Known runtime limits

Tool availability and approvals depend on the Codex environment. A skill can request a capability but cannot grant it.
A skill can be disabled in ~/.codex/config.toml if you need to keep it installed but out of discovery.
OpenAI recommends plugins for reusable distribution. Slashskills v1 intentionally uses direct local Agent Skills instead of plugin packaging.

Official Codex skill docs ↗