Claude consistently uses wrong wildcard syntax in allowed-tools and permission rules

Problem

Claude consistently generates the wrong wildcard syntax when writing allowed-tools in skill frontmatter or permission rules in settings.json. This has been happening for months.

What Claude writes (wrong):

allowed-tools:
  - Bash(~/.claude/skills/my-skill/scripts/my-script.sh *)

"allow": ["Bash(npm *)"]

What it should write (correct):

allowed-tools:
  - Bash(~/.claude/skills/my-skill/scripts/my-script.sh:*)

"allow": ["Bash(npm:*)"]

The correct syntax uses :* (colon-asterisk) for prefix wildcard matching, not * (space-asterisk). The space-asterisk form does not work — it's treated as a literal match looking for a command with a space and an asterisk character.

Impact

• Every time Claude generates a skill with allowed-tools, the rules silently fail to match
• Users get prompted for every command even though the skill explicitly declares permissions
• The user has to manually debug and fix the syntax every time
• This undermines the entire point of allowed-tools in skills

Root cause

The permission rule syntax (Bash(prefix:*)) is documented but Claude's training/system prompt apparently doesn't reinforce it strongly enough. Claude defaults to the more "natural" looking Bash(command *) form, which is wrong.

Suggestion

The system prompt instructions for permission rules should more prominently emphasize the :* colon syntax and explicitly warn against using space-asterisk. Ideally with a "WRONG vs RIGHT" example, similar to what exists for array merging in the update-config skill.

🤖 Generated with Claude Code