Claude Skills explained: a library of instructions, tools and plugins
Claude Skills quickstart: install one skill without replacing your toolbox
A cautious first run for Claude Code or Codex on a disposable workspace
What you will learn
- Choose the native route
- Understand the replacement step
- Test the task, not the badge
Before you start
- A disposable project
- A pinned repository revision
- One supported agent client
Turn a catalogue entry into a measured, reversible team decision
Key takeaways
- The Codex script defaults to installing all skills.
- A same-name destination is replaced, not merged.
- Windows mirror symlinks need special attention.
Choose the native route
For Claude Code, the README documents adding its marketplace and installing an individual plugin. For Codex, the repository offers `scripts/codex-install.sh --skill <name>` and a separate Agent Skills CLI route. These paths do not have identical update and file-copy behavior.
Start with one named skill rather than the script default, which installs all. For a low-risk trial, clone the pinned revision, inspect the chosen source directory, and run the installer with `--dry-run` before a real copy.
Understand the replacement step
The Codex installer resolves its mirrored source, checks for SKILL.md, then removes an existing destination with the same name before copying the replacement with symlinks followed. A dry run reports the destination but does not change it.
Back up a customized destination first. The script does not merge local edits. On Windows, a default Git clone may turn the mirror symlinks into one-line pointer files; the installation guide recommends symlink-aware cloning or copying from real domain directories.
Test the task, not the badge
Open the installed SKILL.md, check which scripts and references came with it, and ask the agent to perform one small task on synthetic files. Compare the result against the same task without that skill.
The article does not claim that a command shown here was executed. Do not run a shell pipe or installer from an unpinned URL simply because it appears in the README; pin, inspect and scope it first.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Pin and inspect the repository revision and chosen SKILL.md.
- 2
Run the single-skill installer with `--dry-run` and review the destination.
- 3
Back up any same-name skill, then test on a disposable task.
Copy-ready example
git clone -c core.symlinks=true https://github.com/alirezarezvani/claude-skills.git
cd claude-skills
git checkout 19392f7a08264ed00486a251f5b2098321771f94
./scripts/codex-install.sh --skill agent-harness --dry-runFrequently asked questions
Can I safely run the Codex installer twice over a customized skill?
Not without preserving that customization: the pinned script removes the existing same-name destination before copying.
Does a successful copy prove the skill works?
No. Confirm that SKILL.md and companion files are real content, then run a small reviewed task.
Sources
- Claude Skills / README.mdSource checked 2026-10-04
- Claude Skills / INSTALLATION.mdSource checked 2026-10-04
- Claude Skills / scripts/codex-install.shSource checked 2026-10-04
- Claude Skills / engineering/agent-harness/skills/agent-harness/SKILL.mdSource checked 2026-10-04