Stop Paying the Marketing Tax: Auditing Your Claude Code Plugin Context

I was burning ~550 tokens per session on marketing plugins I never installed intentionally and didn't need. A 10-minute audit cut my Claude Code baseline context from 33k to 30k tokens, recovered ~3,000 tokens of free headroom, and eliminated noise that was auto-triggering the wrong skills mid-session. Here's exactly how to do it.

Background: What Is the Context Window and Why Should You Care?

If you're using Claude Code for serious development work (large codebases, multi-agent orchestration, long conversion sessions), the context window is your most precious resource. If you're earlier in that journey and wondering what "serious development work" looks like in practice, Nine Months in the Trenches of Agentic Development covers how a production agentic project evolves to the point where a 3,000-token audit actually matters.

Claude Code (running Opus 4.6) has a 200k token context window. That sounds enormous until you realize your baseline overhead is already consuming a significant chunk of it before you write a single line of code.

That baseline includes:

• System prompt: Claude's built-in instructions
• System tools: built-in tool definitions (bash, file ops, etc.)
• MCP tools: any Model Context Protocol servers you've connected
• Custom agents: your /agents definitions
• Memory files: your CLAUDE.md files
• Skills: your /skills definitions
• Plugins: third-party skill bundles installed into Claude Code

Every token of overhead is a token that isn't available for your actual code, conversation history, and file context. On a large project like mine (195k+ lines of legacy VB.NET being converted to ASP.NET Core Razor Pages), this matters.

Step 1: Run the Audit

The first thing to do is see exactly what's consuming your context. In Claude Code, run:

/context

This gives you a detailed breakdown by category. Mine looked like this:

claude-opus-4-6 · 33k/200k tokens (16%)

System prompt:    3.6k tokens  (1.8%)
System tools:    18.4k tokens  (9.2%)
MCP tools:         365 tokens  (0.2%)
Custom agents:     2.0k tokens  (1.0%)
Memory files:      7.4k tokens  (3.7%)
Skills:            1.1k tokens  (0.5%)
Messages:            8 tokens  (0.0%)
Free space:       134k          (67.1%)
Autocompact buffer: 33k         (16.5%)

At first glance, 16% used looks fine. The surprise was in the Plugin section further down the output:

Plugin
└ performance-analytics: 77 tokens
└ content-creation: 76 tokens
└ campaign-planning: 75 tokens
└ competitive-analysis: 73 tokens
└ brand-voice: 68 tokens
└ marketing:seo-audit: 36 tokens
└ marketing:draft-content: 31 tokens
└ marketing:performance-report: 31 tokens
└ marketing:campaign-plan: 30 tokens
└ marketing:email-sequence: 30 tokens
└ marketing:brand-review: 25 tokens
└ marketing:competitive-brief: 25 tokens

12 marketing skills. Installed. In my .NET development environment.

Step 2: Identify the Culprits

I didn't install these manually. After some digging, I found they came from Anthropic's own knowledge-work-plugins/marketing package, a plugin primarily designed for Cowork (Anthropic's agentic desktop application) that also installs into Claude Code.

Here's what each group was actually doing:

The 5 Auto-Triggering Skills

These weren't just sitting dormant. They were auto-triggered. Claude Code scans skill descriptions at session start and loads the ones that appear relevant. During a session where I'm discussing things like "application architecture" or "content strategy for screens," these marketing skills were potentially activating alongside my development tools.

The 7 Slash Commands

The marketing: prefixed skills added explicit commands: /seo-audit, /draft-content, /performance-report, /campaign-plan, /email-sequence, /brand-review, /competitive-brief.

None of these have any place in a Razor Pages migration project.

Step 3: Understand the Real Cost

Raw token count understates the problem. There are two costs:

1. Baseline token burn. Every session starts with ~550 tokens consumed by irrelevant skills. On a long agentic run with context compaction, every token of unnecessary overhead compounds.

2. Wrong skill activation. Skills auto-activate based on description matching. Marketing skills with keywords like "content," "performance," "analysis," and "strategy" have meaningful overlap with software development vocabulary. There's a real risk of Claude reaching for a marketing skill when you meant something technical.

Step 4: Find the Plugin Name

Before uninstalling, confirm the exact registered name:

/plugin list

This shows all installed plugins with their exact names. Look for anything with "marketing" in the name.

Step 5: Uninstall

One command:

/plugin uninstall marketing

If that returns "not found," try the common variants:

/plugin uninstall marketing-skills
/plugin uninstall marketing-suite

Only one will succeed. You're just trying each likely name until you find it.

If the CLI uninstall doesn't work, fall back to manual removal:

# Windows
C:\Users\<you>\.claude\plugins\

# WSL or Mac/Linux
~/.claude/plugins/

Delete the plugin folder directly, then restart Claude Code.

Step 6: Verify

Run /context again immediately after:

/context

The Plugin section should be completely gone from the output. You should also see a small token reduction in the totals.

Results

Here's the before and after:

The cleanup also surfaced several user-level agents that weren't relevant to my current project (api-architect, javascript-expert, date-checker, git-workflow), which I removed at the same time. That accounts for the larger-than-expected agent token reduction.

What Still Has Room to Shrink

After this cleanup, my biggest remaining lever is the project CLAUDE.md at 7.3k tokens. That's the single largest controllable line item, larger than all my custom agents and skills combined.

Strategies I'm evaluating for CLAUDE.md optimization:

• Archive completed screen context: Notes about screens already converted don't need to be in every session's context
• Move reference material to Skills: Static lookup content (naming conventions, component patterns) belongs in skills that load on demand, not permanently in CLAUDE.md
• Lean hierarchy: Break the monolithic CLAUDE.md into a lean root file that references more specific files, letting Claude Code's progressive disclosure load only what's needed

If I can get CLAUDE.md from 7.3k to ~4k, my total baseline drops from 30k to ~27k, which opens up significantly more headroom for complex multi-file conversion sessions.

The Broader Lesson

Claude Code's plugin and skills ecosystem is maturing fast. With community marketplaces, Anthropic's own knowledge-work plugins, and third-party tool suites all installing into the same ~/.claude/ directory, context bloat from unintended installs is a real problem, especially if you use Claude in multiple contexts (development, Cowork, etc.).

Get into the habit of running /context at the start of new projects and periodically during long engagements. Think of it the same way you'd think about reviewing your npm node_modules or NuGet packages. Things accumulate, and periodic audits keep the environment clean. The principle here is the same one I described in Most of the Code AI Learned From Is Garbage: what's in context shapes what the model produces. Garbage in, garbage out — and unnecessary plugin context is the context window equivalent of dependency bloat.

Your context window is your workspace. Keep it clean.

Next in this series: We took the CLAUDE.md from 7.3k tokens to under 2k by rearchitecting from a monolith to a skills-based system that cut always-loaded context by 82%. Read Part 2 →