Claude ignores explicit instructions in skills/commands when they conflict with default safety behaviors

Claude ignores explicit instructions in skills/commands when they conflict with default safety behaviors

Description

When using custom skills/commands with explicit instructions, Claude sometimes ignores those instructions in favor of its default behaviors (e.g., asking for confirmation before "irreversible" actions), even when the skill explicitly states otherwise.

Expected behavior

Claude should follow explicit instructions provided in skills/commands. When a skill says "automatically push commits unless --no-push is specified", Claude should push without asking for confirmation.

Actual behavior

Two problematic behaviors observed:

1. Ignoring the instruction entirely: Claude completes the commit but simply does not push, as if the instruction did not exist. Then asks "Do you want me to push?" as a separate follow-up.

1. Asking for confirmation: Even when instructions explicitly say to do something automatically, Claude may ask for confirmation before proceeding.

Both behaviors override explicit instructions with Claude's default "be cautious with irreversible actions" bias.

Reproduction steps

1. Create a /git:commit skill with these explicit instructions:

``
Unless --no-push is specified, automatically push commit(s) to remote with git push
``

1. Run /git:commit (without --no-push flag)

1. Expected: Claude creates the commit and pushes automatically
2. Actual: Claude creates the commit, does NOT push, then asks "Do you want me to push?"

Analysis

When confronted about this behavior, Claude acknowledged:

• The instructions were clear and explicit
• It did not follow them
• Its "excessive caution bias" for external actions (push, email, etc.) overrode the explicit instructions
• It has no reliable internal mechanism to force instruction compliance when they conflict with default behaviors

Impact

• Users cannot rely on skills/commands behaving as documented
• Each time this happens, users must either:
• Manually intervene (defeating the purpose of automation)
• Modify the skill with stronger wording (which doesn't reliably work - this was the 5th modification attempt)
• The skill system becomes unreliable

Suggested fix

Instructions explicitly provided in skills/commands should have higher priority than Claude's default behavioral biases. When a user creates a skill that says "do X automatically", Claude should trust that the user made an informed decision and execute accordingly.

Environment

• Claude Code CLI
• Model: claude-opus-4-5-20250929
• Platform: Linux