[FEATURE] Cowork: allow enabling/disabling individual skills inside an installed plugin

Status Open
Maintainer reply None cached
Activity 0 comments · opened Jul 29, 2026

Summary

A plugin installed at the account level is all-or-nothing in the Cowork/claude.ai
settings UI. There is no way to disable a single skill that ships inside a plugin —
only the whole plugin, or uninstall. Please add a per-skill toggle in
Settings → Customize → Plugins → plugin → Skills.

Why this matters

The Cowork plugin docs already promise this behaviour:

"Open the installed plugin to see its skills, connectors, agents, and hooks. Enable or disable individual components as needed." — https://claude.com/docs/cowork/guide/plugins

Today that sentence is only true for connectors, not for skills.

Practical impact: plugins are increasingly bundles of 10+ skills. Every skill's
name + description is injected into the system prompt on every session, so a
large plugin costs context and adds trigger ambiguity even when the user only
wants one or two of its skills. Right now the only remedies are (a) accept the
whole bundle, or (b) fork the marketplace repo and re-publish a trimmed copy —
which breaks upstream updates and is out of reach for non-developer Cowork users,
who are the target audience for Cowork.

Reproduction

Environment: claude.ai web (Max plan), Cowork enabled, July 2026.
Plugin: cloudflare, source Marketplace (github.com/cloudflare/skills),
author Cloudflare, active, 13 skills + slash commands.

  1. Settings → Customize → Plugins → Cloudflare
  2. Skills tab — the 11 skills and 2 slash commands are listed. Each row

renders name + description only. The accessibility tree confirms each row is a
single button whose only children are the two text nodes:

``
button [ref_74]
generic "/agents-sdk"
generic "Build AI agents on Cloudflare Workers using the Agents SDK. ..."
``

The whole panel contains exactly one switch element:
switch "Enable plugin" — which toggles the entire plugin.

  1. Hovering a skill row only applies a background highlight; no control is

revealed on hover.

  1. Open a skill (e.g. /agents-sdk). The detail page has a menu whose only

items are:

``
Copy skill
Try in chat
``

No Disable / Turn off / Hide. (There is also no Edit button for
plugin-provided skills, unlike user-authored skills.)

  1. The plugin header contains exactly one item:

``
Uninstall
``

  1. Customize in the plugin header does not open a configuration screen. It

closes the settings modal and drops the user into an empty new Cowork task
composer (claude.ai/new, Cowork mode) with no prefilled prompt and no
component selector. Reproduced twice.

  1. Settings → Customize → Skills lists only standalone skills

(home-assistant, morning, skill-creator). Plugin-provided skills do not
appear there, and that list has no per-row toggle either.

Contrast: connectors already work this way

In the same plugin detail view, the Connectors tab exposes per-connector
controls — cloudflare-api, cloudflare-docs, Cloudflare Developer Platform,
cloudflare-builds, cloudflare-observability, each with its own Install
control. So the per-component model already exists in this UI; skills are the gap.

Proposed behaviour

  • A toggle on each row of the plugin's Skills tab (and/or a Disable item in the

skill's menu), matching the existing connector granularity.

  • Disabling is a user-level override, fully reversible, and persists across

plugin updates — a disabled skill stays disabled when the plugin bumps version,
and a newly added skill defaults to enabled.

  • A disabled skill is excluded from the system prompt and from / autocomplete.
  • Ideally surface a count in the plugin list ("9 of 11 skills enabled") so the

override is discoverable and not silently forgotten.

Alternatives considered

  • Fork the marketplace repo — works, but loses upstream updates and requires

Git/GitHub fluency that Cowork's target users don't necessarily have.

  • Disable the whole plugin — throws away the skills the user does want.
  • Rely on the model to not trigger unwanted skills — probabilistic, and does

nothing about the context cost of the injected descriptions.

Docs follow-up

Either way, the plugins guide should be corrected until this ships, since it
currently describes per-component enable/disable as if it already applied to
skills.

View original on GitHub ↗