Claude Code got dumber. Clean up your skills, one project at a time.

Claude Code, installed and signed in. Run claude update first for a recent version. /skill-doctor needs v2.1.252 or later.

1. 01

   Open Claude Code inside the project you want to clean up. When you close /skills, it saves your choices to .claude/settings.local.json in the folder where Claude Code started. Start it anywhere else and the choices land on the wrong folder.

   cd path/to/your-project claude

   Your part: open a terminal, move into the project folder, and start Claude Code there. Each project keeps its own file, so repeat the guide in every project you want lean.

2. 02

   See what your skills cost before you change anything. Every skill that is on puts its name and description into the skill listing, which Claude reads on every turn. The full skill loads only when it is used. The listing gets 1% of the model's context window, and when it overflows, Claude Code drops descriptions, starting with the skills you invoke least. A skill that loses its description gives Claude less to match your request against. In the setup shown in the video, with 252 skills installed, Jason watched skills that clearly fit the job fail to fire.

   /context /skill-doctor

   Your part: run both and read. /context shows a Skills row with the listing's size. /skill-doctor shows what each skill costs and how often it gets used, and flags skills you have never invoked. On an older version, /doctor estimates the listing cost. The only overflow warning goes to the debug log, so you will not see it in the chat.

3. 03

   Let Claude draft the cleanup. This prompt has Claude read the project, list your skills with their current states, and suggest a state for each one with a reason. It stops and changes nothing until you answer.

   Paste into Claude Code

   Copy prompt

   Audit my Claude Code skills for this project.
   1. Read this project first: the README, the top-level folders and the main config files, so you know what work happens here.
   2. List every skill I have: the ones you can see in this session, plus every folder in ~/.claude/skills and .claude/skills. Leave out plugin skills, whose names contain a colon (like vercel:deploy). I manage those in /plugin.
   3. Read skillOverrides in .claude/settings.local.json in this folder and in ~/.claude/settings.json, so you know each skill's current state.
   4. Show me a table: skill, current state, suggested state (on, name-only, user-only or off), and a one-line reason. My rule: relevant and helpful to this project means on. Flag any skills that do the same job, like several writing skills, and suggest keeping one on.
   5. Stop there and change nothing until I answer.
   6. When I say yes, or give you changes, merge the states into skillOverrides in .claude/settings.local.json in this folder. Write user-only as "user-invocable-only". Keep every other key and entry in that file, and do not touch ~/.claude/settings.json.
   7. Then tell me to run /skills to check the result.

   Your part: read the table and change any row you disagree with. Then say yes, approve the file edit when Claude Code asks, and open /skills to check the result. Read every row before you agree.

4. 04

   Check and adjust in /skills. Type to filter, press t to sort by token count, and press Space or Enter to cycle a skill through four states. On: Claude sees the name and description, and /name is in the menu. Use it for skills Claude should reach for on its own in this project. Name-only: Claude sees the name alone, /name stays in the menu, and the listing gets room back. Use it when the name says what the skill does. User-only: Claude does not see it, and you can still type /name. Use it for skills you want only when you call them. Off: hidden from Claude and from the / menu. Use it for skills this project never needs. A skill with no entry counts as on.

   /skills t sort by tokens Space / Enter cycle the state Esc save and close

   Your part: decide each one, then press Esc to save. How Jason, who made the video, decides: "Is it relevant and helpful to this project? If yes, on. If no, off." Name-only and user-only cover the skills in between. The ~N tok beside each skill estimates what its listing entry costs.

5. 05

   Break up stacks. Keep one skill per job on in a project. When two skills promise the same outcome, both descriptions compete for Claude's attention on every turn. Claude Code's docs put it plainly: "If descriptions are vague or overlap, Claude may load the wrong skill or miss one that would help."

   Warning, from Jason's own setup: he had humanizer, copywriting and the-writer all on in one project. They stacked and collided, and in his experience the writing came out worse. Pick the one that fits this project, leave it on, and set the others to user-only so you can still call them by name.

6. 06

   Turn off plugins this project does not use. Plugin skills show a lock in /skills and cannot be cycled there, so you switch the whole plugin instead. In /plugin, open the Installed tab, where a Not used recently group points at likely candidates, and press Enter on a plugin to enable, disable or uninstall it. To turn a plugin off for this project only, run the command below from the project folder. It writes to the same .claude/settings.local.json.

   /plugin > Installed > Not used recently

   Run in a terminal, from the project folder

   Copy prompt

   claude plugin disable <plugin>@<marketplace> --scope local

   Your part: swap in the plugin and marketplace names as /plugin shows them. Leave on anything this project uses. Disable keeps a plugin installed, so you can turn it back on later. Uninstall removes it.

7. 07

   Set a lean default for everywhere else. /skills saves to the project folder, so the global default goes into ~/.claude/settings.json by hand. Add a skillOverrides block and merge it into the file you already have, keeping every other key. Both files apply together, skill by skill. When both set the same skill, Claude Code's documented settings order puts the project's local file first, so a project can turn back on a skill that is off everywhere else.

   Add by hand to ~/.claude/settings.json

   Copy prompt

   {
     "skillOverrides": {
       "some-skill": "off",
       "another-skill": "name-only",
       "a-third-skill": "user-invocable-only"
     }
   }

   Your part: swap in your own skill names and states, and keep the file valid JSON. On Windows the file is C:\Users\<you>\.claude\settings.json. Then open /skills in a project and confirm each skill shows the state you expect.

8. 08

   Optional: the other two fixes from the reel. ponytail pushes Claude toward the simplest solution that works, the fix for reinventing the wheel on every small change. humanizer strips the patterns that make text read as machine-written. ponytail installs as a plugin, so it shows a lock in /skills.

   ponytail on GitHub

   Install ponytail, then humanizer

   Copy prompt

   In Claude Code, send these as two separate messages:
   /plugin marketplace add DietrichGebert/ponytail
   /plugin install ponytail@ponytail
   
   In a terminal:
   npx skills add blader/humanizer --global

   Your part: send the ponytail lines one at a time. Leave off --global to install humanizer for this project only. The npx command needs Node.js. Prefer humanizer as a plugin? Send /plugin marketplace add blader/humanizer, then /plugin install humanizer@humanizer. Source: github.com/blader/humanizer. Then give each one the state it earns in this project.