[Bug] Skills not listed in natural language queries and inaccurate token counting for skill context
Bug Description
Feedback: Skills vs Commands confusion and inaccurate token counting
Issue 1: Skills are NOT listed when asking via prompt - only commands are
What happened:
When I ask Claude via natural language to "list available skills" (e.g., "Liste moi tes skills et leurs descriptions"), it responds with COMMANDS only (e.g. ~/.claude/plugins/.../commands/).
The problem:
Skills (e.g. ~/.claude/plugins/.../skills/) are completely absent from the response. This is extremely confusing because:
- Skills and Commands are different concepts in Claude Code
- When a user asks for "skills", they expect to see skills, not commands
- This makes skills essentially invisible through natural conversation
Example:
- Prompt: "List your skills"
- Claude returns: /commit, /review-pr, /create-agent (commands)
- Missing: skill-writing, git-commiting, etc. (actual skills)
---
Issue 2: Highly inaccurate token counting - and concerning implications
What happened:
The /skills command displays severely underestimated token counts:
- git-branch-taskforce · ~22 tokens (actually 300+ tokens)
- archive-done · ~21 tokens (actually 300+ tokens)
- git:git-commit · ~8 tokens (actually 300+ tokens)
- cc-toolkit:claude-md-writing · ~13 tokens (actually 300+ tokens)
Token counts are off by 10-20x.
More concerning:
The /context command shows the same incorrect token counts as /skills. This raises a serious question: Are skill descriptions actually being read by Claude?
Why I'm worried:
- My skills are not being invoked proactively nearly as often as they should be
- The inaccurate token accounting makes me suspect the skill descriptions aren't actually loaded into context
- This would explain why Claude doesn't recognize when to use them
- Note: I'm 100% certain my skill frontmatter is correct, so this is not a configuration issue on my end
Impact:
- Skills become much less useful if they don't activate automatically
- Users lose trust in the system when core features don't wos documented
- Inaccurate token counts prevent proper context budget management
---
Environment Info
- Platform: darwin
- Terminal: cursor
- Version: 2.1.3
- Feedback ID: 29fe192d-bcf2-462f-bb9c-82cac48c5ac7
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗