Which skills is Claude Code carrying? The ones you didn’t install, and what survives /compact
Open a fresh Claude Code session and type /skills. The list has skills you never installed: some ship with Claude Code, and if you sign in with a claude.ai account, the skills turned on there come along too, pdf and xlsx always. Every skill Claude can pick by itself adds to the context on every turn, used or not. And a skill you did use isn’t safe either: after compaction, only its first 5,000 tokens come back. Here’s how to see what’s loaded, turn off what you don’t use, and check that the skill you need is still all there.
- Where they come from
- See what’s loaded
- What they cost
- Turn them off
- After /compact
- Check a skill is still there
- Questions
Where the skills you didn’t install come from
| Source | What you get | When |
|---|---|---|
| Bundled with Claude Code | /code-review, /debug, /loop, /batch, /simplify, /run, /verify, /claude-api, /dataviz, /update-config, /fewer-permission-prompts and a few more | Every session |
| Your claude.ai account | The skills turned on in your claude.ai settings, plus some of Anthropic’s own, such as pdf and xlsx, which always sync | When you’re signed in with that account, and in cloud and Cowork sessions. Not when you use an API key or Bedrock |
| Plugins you enabled | Whatever skills each plugin carries, named /plugin-name:skill-name | Wherever the plugin is on |
Bundled skills are prompts: Claude reads the instructions and follows them, the same way it follows yours. Commands such as /compact, /context and /init look the same in the / menu but aren’t skills. Synced skills download into ~/.claude/skills/synced/, and the sync doesn’t slow down startup.
See what’s loaded
/skillslists every skill in the session. The ones from your account sit in their own group, claude.ai sync./contextshows how much of the context the skill list takes, in the Skills row, with synced skills grouped the same way./skill-doctorshows what each skill costs and how often you use it, and flags the ones you’ve never run. It leaves out bundled skills, so use it to clean up your own and your plugins’. It needs Claude Code 2.1.252 or later./statuslists skill files that are broken and won’t load, along with broken agent, command and output style files. That list arrived in version 2.1.327.
What a skill costs before you use it
- Name and description, on every turn. Every skill Claude can pick by itself sends its name and description with every request, whether Claude ever uses it or not. A description is cut at 1,536 characters, so one skill costs at most a few hundred tokens a turn. A dozen you never use add up.
- The body only when used. The instructions load when the skill runs, so a long skill costs nothing until then.
- The list has a budget of 1% of the model’s context window. When there are more descriptions than fit, Claude Code drops the descriptions of the skills you use least. Claude can’t pick a skill whose description it can’t see; you can still run it with
/name.
Turn off the ones you don’t use
| To turn off… | Do this |
|---|---|
| One skill from claude.ai | Turn it off in your claude.ai skill settings. pdf and xlsx always sync, so for those turn the sync off |
| Everything synced from claude.ai | "syncClaudeAiSkills": false in ~/.claude/settings.json or .claude/settings.local.json. In the shared .claude/settings.json it’s ignored. It stops the download and the skills already synced, which move to ~/.claude/skills/.trash/ at the next start |
| All bundled skills | "disableBundledSkills": true. Commands such as /init keep working |
| Any one skill, or just its description | skillOverrides, below |
| Your own skill that you only run by hand | disable-model-invocation: true in its frontmatter: its description leaves the context, and /name still works |
| A plugin’s skills | Disable the plugin. skillOverrides doesn’t touch plugin skills |
skillOverrides maps a skill’s name to one of four values. It changes who sees the skill, not what the skill says:
"on", the default: Claude sees the name and description, and so do you in the/menu."name-only": Claude sees only the name."user-invocable-only": hidden from Claude, still in your/menu."off": hidden from both.
You can also change it in the /skills menu: Space cycles the value, Esc saves it to .claude/settings.local.json.
~/.claude/settings.json
{
"syncClaudeAiSkills": false,
"skillOverrides": {
"dataviz": "off",
"loop": "user-invocable-only"
}
}
Before you turn things off, run /skill-doctor for a week’s worth of real use. Keep what you ran, cut what you didn’t.
What’s left of a skill after /compact
When a skill runs, its full text goes into the conversation and stays there. Compaction replaces the conversation with a summary, and then Claude Code puts the skills you used back, with two limits:
- The first 5,000 tokens of each skill. Anything below that line is gone, so a long skill loses its last steps.
- 25,000 tokens for all of them, filled from the skill you used most recently. If you ran several long skills, the oldest may not come back at all.
It’s easy to miss. Claude carries on with the part it still has, and what’s missing is whatever was at the end, often the final check or the cleanup. So:
- Put what matters most at the top of
SKILL.md, and keep the file short. Move long reference material into separate files in the skill’s folder and point to them from the instructions. - Run the skill again after compaction if you still need all of it. That’s the documentation’s own advice. Then check it with the check word below.
- Compact before you start a skill-heavy task, not in the middle of it (when to compact and when to clear).
Check that a skill is still all there
A check word at the end. Add a line with a rare word to the very end of SKILL.md. It means nothing to Claude while it works, but if the end of the skill is in context, Claude can quote it.
SKILL.md, last line
Check word: heron-41
Then, any time you’re not sure, ask:
Quote the check word at the end of the release skill, exactly as written. If you can't see it, say so. Don't guess.If Claude can’t quote it, the end of the skill is gone. Run the skill again.
A reminder after every compaction. A SessionStart hook with the matcher compact runs right after compaction, and whatever it prints is added to the context. Use it to tell Claude to reload the skill it was following:
.claude/settings.json
{
"hooks": {
"SessionStart": [
{
"matcher": "compact",
"hooks": [
{
"type": "command",
"command": "echo 'Context was compacted. Rerun any skill you were following.'"
}
]
}
]
}
}
/skill-doctor isn’t this check. It tells you what skills cost over time, not what’s in this conversation right now.
Where Meowdoro fits
We built Meowdoro, a focus timer around the MacBook notch for people who work with Claude Code and Codex. Each session gets a ring that shows how full its context is, and the skill list is part of what fills it from the first message. When a ring gets close to full, compaction is near, and that’s the moment to finish or rerun the skill you’re in. Try it and tell us what’s missing.
Questions
What skills come with Claude Code by default?
Bundled skills such as /code-review, /debug, /loop, /batch, /simplify, /run and /verify. If you sign in with a claude.ai account, the skills turned on there load too, and pdf and xlsx always do. Run /skills to see your list.
Why are there pdf and xlsx skills in my Claude Code?
They sync from your claude.ai account. Some of Anthropic’s skills always sync when you’re signed in with it. Turn the sync off with "syncClaudeAiSkills": false in ~/.claude/settings.json.
Do skills use tokens if I never use them?
Yes, a little. Every skill Claude can pick by itself sends its name and description on every turn. The body loads only when the skill runs. disable-model-invocation: true or skillOverrides takes a skill out of that list.
Why did Claude stop following my skill after /compact?
After compaction Claude Code puts back only the first 5,000 tokens of each skill you used, 25,000 in all, newest first. The end of a long skill, or an older skill, can be gone. Run the skill again with /name, and keep the key instructions near the top of SKILL.md.
How do I check that a skill is loaded?
/skills shows what’s available. To check that a skill you used is still all in the conversation, put a rare check word on its last line and ask Claude to quote it.
Related guides
How skills load, sync, cost context and come back after compaction: the Claude Code documentation on skills, settings and hooks, read on November 18, 2026. What /status lists: the Claude Code changelog, version 2.1.327, read on November 19. The check word and the hook message are examples, not taken from the documentation.