Feature request: enable/disable individual skills via settings.json and session flags

[Brenndy]

Currently, Claude Code reads all SKILL.md files found under .claude/skills/ and activates them based on their description: trigger. There is no way to toggle individual skills on or off without physically removing or renaming the file.

Proposed — persistent toggle in settings.json

{
  "disabledSkills": ["worktree-setup"]
}

This would allow teams to ship a set of skills in a repo but let individual developers or CI environments opt out of specific ones without modifying tracked files.

Proposed — session-scoped toggle via CLI flag

claude --disable-skill worktree-setup
claude --enable-skill worktree-setup  # re-enable one disabled in settings.json

Or as a slash command mid-session:

/skill disable worktree-setup
/skill enable worktree-setup
/skill list

Use cases

• A repo has 5 skills installed. A developer wants to disable one that conflicts with their workflow without removing it from the repo or touching .claude/skills/
• CI pipeline wants to run with a minimal skill set without modifying tracked files
• A developer temporarily disables a skill for one session while debugging, without affecting teammates

[cdelgado70]

Strong agreement with the need. The current behavior — every installed skill loads in every project — creates a perverse install-time incentive: every speculative skill costs available_skills budget and selection-attention forever, so users self-censor what they install. With finer-grained enable/disable (per-project or per-task), users could install many more skills and only pay for the ones relevant to a given session.

Related ticket on the SkillKit side: vercel-labs/skills#634 — same need, different surface. Neither has shipped.

As a workaround in the meantime, a search-based discovery layer can sidestep the native listing entirely: index your skills out-of-band, inject the relevant section per-prompt, and move ~/.claude/skills/ to a disabled location so Claude Code doesn't enumerate it. I measured ~2,007 tokens/turn savings on a 95-skill installation. Writeup if interesting: https://cdelgado70.github.io/2026/05/06/skills-and-the-discovery-ceiling.html