Feature Request: Global option to disable skill auto-triggering at project level

[xuxinpeng]

Preflight Checklist

• I have searched existing requests and this feature hasn't been requested yet
• This is a single feature request (not multiple features)

Problem Statement

Skills (slash commands) can be automatically triggered by Claude based on context-defined trigger conditions injected at the system level. There is no official way to globally disable this auto-triggering behavior at the project level.

The only available option is per-skill frontmatter (disable-model-invocation: true), but this requires modifying each skill individually and does not work for built-in/system-provided skills.

Proposed Solution

Add a project-level configuration option in .claude/settings.json to globally disable skill auto-triggering:

{
  "skills": {
    "autoTrigger": false
  }
}

When set to false, skills should only be invocable by the user explicitly typing /skill-name, and Claude should not automatically trigger any skill based on context.

Alternative Solutions

1. Adding instructions in CLAUDE.md to tell Claude not to auto-trigger skills — this is a soft constraint and not reliably enforced at the system level.

2. Adding disable-model-invocation: true to each skill's frontmatter — only works per-skill, not globally, and cannot be applied to built-in/system-provided skills.

Priority

Medium - Would be very helpful

Feature Category

Configuration and settings

Use Case Example

A team wants strict control over skill usage in their project. They add {"skills": {"autoTrigger": false}} to .claude/settings.json. After this, Claude no longer auto-invokes skills based on context (e.g., it won't auto-trigger the "claude-developer-platform" skill when it sees an anthropic import). Skills can still be used by manually typing /skill-name.

Additional Context

No response

[yurukusa]

A PreToolUse hook on Skill can block auto-triggered skills:

INPUT=$(cat)
SKILL=$(echo "$INPUT" | jq -r '.tool_input.skill // empty')
USER_INVOKED="/tmp/cc-user-invoked-skill"
if grep -qF "$SKILL" "$USER_INVOKED" 2>/dev/null; then
    sed -i "/$SKILL/d" "$USER_INVOKED"
    exit 0
fi
echo "BLOCKED: Skill '$SKILL' was auto-triggered. Use /$SKILL to invoke explicitly." >&2
exit 2

Track user invocations:

PROMPT=$(cat | jq -r '.userPrompt // empty' 2>/dev/null)
if echo "$PROMPT" | grep -qE '^/[a-z]'; then
    SKILL=$(echo "$PROMPT" | grep -oE '^/[a-z-]+' | sed 's/^\///')
    echo "$SKILL" >> /tmp/cc-user-invoked-skill
fi
exit 0

This ensures skills only fire when you type /skill-name, never when the model decides to auto-trigger them.