Claude Skills explained: a library of instructions, tools and plugins
Deploying Claude Skills for a team
Make skill versions, destinations and rollback ownership explicit before a broad rollout
What you will learn
- Choose the distribution boundary
- Package a narrow release
- Plan reversal
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
- Install paths have different update semantics.
- Conversion is a build step, not evidence of parity.
- Rollback requires the prior files.
Choose the distribution boundary
Claude Code users can install individual plugins through the marketplace; Codex users can copy mirrored skill packages; other tools may need converted rules. Select one path per client and record the revision that produced its files.
A shared repository and a per-user home directory have different ownership. Decide whether skills are project-scoped or user-scoped before installation, and avoid letting an unreviewed home-directory update silently change every project.
Package a narrow release
Create a manifest of selected skill names, source revision, content hashes and destination paths. Verify SKILL.md plus required scripts and references after copying. A conversion step can change format, so retain the source package and generated output separately.
The pinned installation docs describe multiple platforms but do not establish that every skill behaves identically in each agent. Run a smoke task in the actual client version and document any unsupported triggers or commands.
Plan reversal
Keep the previous destination as a recoverable snapshot before a replacement. The Codex script deletes a same-name target during install, so a rollback is a restore of files and configuration, not merely changing a version string.
For a team rollout, restrict the first cohort, collect failures and agent traces without secrets, then promote the exact artifact. We did not deploy this library to a team or validate cross-platform parity.
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
Choose project or user scope for each client.
- 2
Record source revision, hashes and destinations.
- 3
Pilot, snapshot, promote and test rollback.
Copy-ready example
release: reviewed-revision
selected_skills: [agent-harness]
destination: scoped-client-directory
verify: [SKILL.md, required-scripts, smoke-task]
rollback: previous-files-snapshotFrequently asked questions
Can a team just track the repository main branch?
That makes each install potentially different. Pin a revision and review the resulting files before promotion.
Does a Codex copy automatically track upstream updates?
No. The pinned installer copies content; later updates require a new reviewed copy.
Sources
- Claude Skills / INSTALLATION.mdSource checked 2026-10-04
- Claude Skills / .claude-plugin/marketplace.jsonSource checked 2026-10-04
- Claude Skills / scripts/codex-install.shSource checked 2026-10-04
- Claude Skills / scripts/convert.shSource checked 2026-10-04
- Claude Skills / scripts/install.shSource checked 2026-10-04