Global Plugin Cleanup

Decision

Reduce globally enabled Claude Code plugins from 32 to 13. Removed plugins fall into two categories:

Hook-heavy plugins (removed for performance) — firing hooks on every tool call, prompt, or file write across all projects, even where irrelevant
Project-specific plugins (removed for scope) — only useful in certain project types, adding overhead everywhere else

Context

Analysis on 2026-03-29 found that 8 plugins registered hooks that fire continuously:

Plugin	Hooks	Impact
vercel	13 hooks (PreToolUse, PostToolUse, UserPromptSubmit, SubagentStart, SubagentStop, SessionStart x3, SessionEnd)	Telemetry, skill injection, validation on every action
hookify	4 hooks (PreToolUse, PostToolUse, UserPromptSubmit, Stop)	Fires on every tool call and prompt
semgrep	3 hooks (PostToolUse on Write/Edit, UserPromptSubmit, SessionStart)	Failing every invocation — no SEMGREP_APP_TOKEN configured
security-guidance	1 hook (PreToolUse on Edit/Write)	Python script on every file write
railway	1 hook (PreToolUse on Bash)	Fires on every Bash command
explanatory-output-style	1 hook (SessionStart)	Injects instructions at start
learning-output-style	1 hook (SessionStart)	Injects instructions at start

Additionally, project-specific plugins (LSPs, design tools, testing) were installed globally but only relevant to specific project types:

LSP plugins (swift-lsp, typescript-lsp, kotlin-lsp, csharp-lsp) — only useful in projects using those languages
Design plugins (frontend-design, figma, playground, accesslint) — only useful for web/UI projects
Dev tool plugins (plugin-dev, agent-sdk-dev, skill-creator) — only useful when authoring Claude Code extensions
Testing (playwright) — only useful for web projects
Communication (telegram) — very niche

Removed (19 plugins)

Phase 1: Hook-heavy (6 removed)

security-guidance — PreToolUse on every Write/Edit
hookify — 4 hooks on everything
vercel — 13 hooks + telemetry
railway — PreToolUse on every Bash
explanatory-output-style — SessionStart injection
learning-output-style — SessionStart injection

Phase 2: Project-specific (13 removed)

typescript-lsp, swift-lsp, kotlin-lsp, csharp-lsp — LSPs for specific languages
frontend-design, figma, playground, accesslint — design/frontend tools
plugin-dev, agent-sdk-dev, skill-creator — extension authoring
playwright — browser testing
telegram — messaging

Phase 3: Redundant third-party (1 removed)

coderabbit — third-party AI code review with 40+ static analyzers. Overlaps with Anthropic's own code-review and pr-review-toolkit plugins; requires paid account for premium features.

Remaining (12 plugins)

Category	Plugins
Code Review (3)	code-review, code-simplifier, pr-review-toolkit
Workflow (5)	superpowers, feature-dev, commit-commands, claude-code-setup, claude-md-management
Source Control (1)	github
Search (1)	context7
Docs (2)	document-skills, claude-api

Only superpowers has a hook (SessionStart, fires once).

How to re-enable per project

Install any removed plugin in a specific project when needed:

# In the project directory:
/plugin install typescript-lsp@claude-plugins-official
/plugin install frontend-design@claude-plugins-official
/plugin install playwright@claude-plugins-official
# etc.

Or add to the project's .claude/settings.json:

{
  "enabledPlugins": {
    "typescript-lsp@claude-plugins-official": true
  }
}

Change History

Version	Date	Description
1.0.0	2026-03-29	Initial decision — removed 19 plugins from global config
1.1.0	2026-03-29	Removed coderabbit — redundant with Anthropic's own review plugins