Last week I wrote about measuring what my Claude Code setup costs on every request. The short version: 68.5k tokens before I typed a word, and 220 of my 283 skills listed with no description at all.

The obvious follow-up question was: what do you actually do about it? Uninstalling everything you didn't use this week is a blunt answer. It turns out Claude Code has settings for exactly this, and I hadn't heard of any of them.

Why skills lose their descriptions

Every session starts with a listing of your skills: a name and a description for each. Claude reads the descriptions to decide when a skill applies.

That listing has a budget. According to the docs it scales at 1% of the model's context window. On a 1M-token model that came out at 29,997 characters in my transcripts. When the listing overflows, Claude Code keeps every name but drops descriptions, starting with the skills you invoke least.

That matches what I saw. All 13 skills I'd used in the last month kept their descriptions. The 220 that went blank were all skills I hadn't touched. The catch is that a skill you just installed has zero invocations too, so it's first in line to go blank, and Claude can't reach for it on its own because all it sees is a name.

The three settings

skillOverrides

Decides how each skill shows up:

{ "skillOverrides": { "old-report-skill": "user-invocable-only", "deploy": "name-only" } } 
value Claude sees in the / menu
"on" (default) name + description yes
"name-only" name only yes
"user-invocable-only" nothing yes
"off" nothing no

user-invocable-only is the useful one. The skill disappears from what the model sees, but you can still run it yourself with /name. For a skill you haven't used in a month, that's usually what you want.

One catch: plugin skills aren't affected by skillOverrides. For those, the only option is disabling the plugin.

skillListingBudgetFraction

This is the 1% itself. 0.02 doubles the room. It isn't free, though: my listing is about 30k characters, so doubling it can cost up to ~8k more tokens on every request once it fills up. Hiding the skills you don't use is the cheaper way to get room back.

disableClaudeAiConnectors

If you've connected services on claude.ai, Claude Code loads their tool names and instructions in every session. Setting this to true stops Claude Code from fetching them. They stay connected on claude.ai. If you use some connectors but not others, toggle the unused ones off in /mcp instead (that's stored per project).

Doing it from your actual usage

Choosing what to hide by hand isn't fun with 283 skills, so I added it to deadweight, the tool from the last post. It already reads the transcripts in ~/.claude/projects to see what you used in the last 30 days. Version 0.3.0 uses the settings above:

cargo install deadweight deadweight fix --dry-run 
Disable these plugins (settings.json → enabledPlugins: false) ... Hide these skills from Claude (settings.json → skillOverrides: "user-invocable-only"; /name still works) ... Turn off claude.ai connectors in Claude Code (settings.json → disableClaudeAiConnectors: true) ... Saves ~5.3k tokens per request. Undo any time with deadweight undo. 

On my machine that's 17 unused plugins, 30 unused skills and one connector. The previous version moved skill folders into a backup. Now nothing on disk moves: it's a few keys in settings.json, and deadweight undo puts back exactly what was there before, including overrides you had already set yourself. --park-skills keeps the old behaviour if you prefer it.

When descriptions are being dropped, the report also does the budget math for you:

✗ 220 of 283 skills are listed without a description: the skill listing is full (~30.0k characters). Claude sees only their names, so it can't tell when to use them. Make room: hide what you don't use (deadweight fix), or set "skillListingBudgetFraction": 0.02 in settings.json (default 0.01; the extra room costs up to ~8.3k tokens per request). 

What I'd still check by hand

  • Plugins that run hooks. deadweight won't disable them unless you pass --include-hooked, because a hook can do useful work that never shows up as a tool call. On my machine that's a memory plugin I'd rather keep.
  • Skills you use rarely. Claude Code deletes transcripts after 30 days by default (cleanupPeriodDays), so a skill you use once a quarter looks unused. --keep NAME protects it.
  • Servers you never signed in to. I had 36 of them still listing authenticate tools in every session. They all came from plugins, so disabling the plugin is what removes them.

A lot of this was worked out by people measuring their own listings in anthropics/claude-code#81081. It's worth a read if you want the details of the drop order.

If deadweight gets something wrong on your setup, run deadweight --json and open an issue. Those reports are the most useful thing anyone can send me.