Enabling and Managing Skills Across Claude Code, Cowork, and Claude.ai

A Skill built once is meant to work the same way everywhere - the instructions inside SKILL.md don't change depending on whether you're in Claude Code, Cowork, or the claude.ai web app.

What does differ, surface to surface, is where a Skill lives and how you turn it on.

This page walks through enabling and organizing Skills on each of the three, and what to check when a Skill behaves differently in one place than another.

Recipe

• Identify which surfaces you actually use day to day - Claude Code for a codebase, Cowork for team workspace tasks, claude.ai for everyday chat - since each has its own enablement step.
• In Claude Code, place the Skill folder in the project or user-level Skills location the tool expects, so it's picked up for that repository or across all your projects.
• In Cowork, add the Skill to the team workspace's Skills area, where it becomes available to conversations inside that workspace.
• In claude.ai, add the Skill through your account's Skills settings, where it becomes available across your personal conversations.
• After enabling in any surface, test with a real, natural request rather than a generic message, and confirm the output matches what the Skill's instructions describe.
• Revisit each surface's list periodically - disabling Skills you no longer use keeps discovery cleaner and avoids overlapping descriptions competing for the same tasks.

Working Example

Enabling the same Skill across all three surfaces looks like this in practice.

Skill: "code-review-checklist"
Goal: available in Claude Code (this repo), Cowork (team workspace),
and claude.ai (personal account).
 
1. Claude Code: add the Skill folder to the project's Skills location
   so it applies whenever you're working in this repository.
2. Cowork: add the same Skill folder to the team workspace's Skills
   area so teammates get it too, not just you.
3. claude.ai: add the Skill through personal account settings so it's
   available in ordinary chat, outside any workspace or repository.
4. Test each surface separately: ask for a code review in a way you'd
   naturally phrase it, and confirm the checklist format shows up.

Each step is a separate enablement action, even though the underlying SKILL.md file is identical across all three - nothing about the Skill's content needs to be rewritten per surface.

The test step matters more than it looks like: a Skill enabled correctly on one surface but never tested there can silently fail to fire if that surface's enablement didn't fully take, and the only way to catch that is a real request.

Deep Dive

The reason enabling differs by surface, while behavior once loaded doesn't, comes down to what each surface controls.

Claude Code is typically scoped to a project or codebase, so its Skills location is often tied to that project, with an option for Skills that apply across all your projects.

Cowork is built around shared team workspaces, so its Skills area is naturally scoped to a workspace - a Skill added there becomes available to everyone working in that space, not just the person who added it.

claude.ai is scoped to your personal account, so Skills added there follow you across your own conversations, independent of any team or repository.

None of that changes how discovery works once a Skill is enabled - Claude still checks the same description field against the same kind of incoming task, using the same matching logic, regardless of which surface the conversation is happening on.

What can genuinely differ across surfaces is tool access: allowed-tools in a Skill's frontmatter interacts with whatever tools that surface exposes in the first place, so a Skill that references a tool only available in Claude Code won't be able to use it inside claude.ai's chat interface, even if the Skill itself is enabled there.

Managing a growing Skill library across three surfaces is mostly a matter of keeping one canonical version of each Skill and deciding deliberately where it needs to be enabled, rather than letting slightly different copies drift apart on different surfaces.

Gotchas

• Enabled in one surface, tested in another. Confirming a Skill works in claude.ai doesn't guarantee it behaves the same way in Claude Code if it was never actually enabled there.
• Team-scoped Skills mistaken for personal ones. A Skill added to a Cowork workspace is available to that workspace's conversations, not automatically to your personal claude.ai account.
• Tool references that don't exist on every surface. A Skill written with a Claude Code-specific tool in mind may load correctly elsewhere but be unable to actually use that capability.
• Stale copies drifting apart. Editing a Skill's instructions on one surface without updating the same Skill elsewhere leads to inconsistent behavior across surfaces over time.
• Assuming re-enabling is needed after every edit. Most edits to an already-enabled Skill's SKILL.md take effect on save, without a separate re-enable step - but it's worth confirming on each surface rather than assuming.

Alternatives

• Custom Instructions instead of a Skill, when the goal is a standing preference that should apply to every conversation on a surface, not a task-triggered recipe.
• A Project's persistent context, when the goal is information tied to one ongoing workspace rather than a repeatable task you want available everywhere.
• Keeping a Skill enabled on a single surface only, when a task genuinely only ever comes up in one context - for example, a Claude Code-specific workflow with no equivalent in ordinary chat.

FAQs

Related

• Agent Skills Basics - creating your first Skill before deciding where to enable it
• Skills vs Projects vs Custom Instructions: What's the Difference - when a Skill is the right tool versus a Project or Custom Instructions
• Anatomy of a SKILL.md File - the allowed-tools field referenced in the Deep Dive above
• Signs Your Skill Isn't Being Discovered by Claude - diagnosing a Skill that seems enabled but isn't firing

Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.