Skills don't auto-activate in spawned agents (via Task tool)

Summary

Skills do not auto-activate in agents spawned via the Task tool, even when prompts match skill trigger descriptions. Explicit instruction is required for spawned agents to use the Skill tool.

Expected Behavior

Based on documentation, skills should auto-activate when:

• Skill description matches the task context
• Prompt contains trigger keywords from skill description

Actual Behavior

• Main session: Skills can auto-activate (higher rate, has rule context)
• Spawned agents: 0% auto-activation rate without explicit instruction

Reproduction

1. Create a skill with clear trigger conditions in description
2. Spawn an agent via Task tool with a prompt matching those conditions
3. Observe: Agent does not invoke Skill tool

Evidence

Rigorous testing conducted (December 2025):

| Test Type | Auto-activation Rate |
|-----------|---------------------|
| Spawned agent - no instruction | 0% (0/20 tests) |
| Spawned agent - explicit instruction | 100% (5/5 tests) |

Root Cause Analysis

1. Rules in .claude/rules/ are session context only - they don't propagate to spawned agents
2. Spawned agents start fresh with only their prompt and agent definition
3. The skills: field in agent frontmatter doesn't auto-load skill content (documented elsewhere)

Current Workarounds

Users must explicitly instruct spawned agents to be skill-aware:

1. In agent .md file:

``markdown
**FIRST:** Use the Skill tool to invoke "using-superpowers", then proceed.
``

1. In Task prompt:

``
Check if ANY skill applies to this task. Use the Skill tool if you find one.
``

Impact

• Reduces effectiveness of skill system in agentic workflows
• Requires manual workarounds for skill-aware agents
• Documentation implies auto-activation should work

Suggested Improvements

1. Propagate skill context: Pass available skill metadata to spawned agents
2. Add skill-aware flag: Allow agent definitions to opt into skill checking
3. Document limitation: Update docs to clarify spawned agent behavior

Environment

• Claude Code CLI
• Tested with Haiku, Sonnet, and general-purpose agents
• December 2025