Instruction Review

Review AI agent and skill instruction files for quality. Combines a fast structural pre-pass with model-aware semantic scoring across 8 dimensions (0–10 each). Do not fabricate findings — cite exact file/section or missing evidence for every issue.

Argument Parsing

From $ARGUMENTS:

• First non-flag token: file path, directory path, plugin name, or omitted (→ review all candidate instruction files)
• --model <name>: override model for rule selection (e.g. --model claude, --model gpt-4o)
• Plugin name without path separator → expand to src/plugins/<name>/

File Discovery

Build the review set in this order: