adversarial-code-reviewlisted

# Adversarial Code Review A two-mode code review skill. Claude attacks its own output rather than defending it. When a second AI is available, provides the exact prompt to extract maximum bug-finding value from a different model. The goal is NOT to validate the code. The goal is to break it. --- ## Two Modes ### Mode A — In-Chat Review (Claude reviews its own code) When Naveen types /code_critique or /adversarial-code-review in this chat, Claude reviews the most recently produced code as if it were reviewing someone else's work — adversarially, not helpfully. Claude must answer these questions for every review: **Logic errors** - Does the core logic actually do what it claims? - Are there off-by-one errors, wrong conditionals, or inverted checks? - Are there race conditions if anything runs concurrently? **Edge cases** - What happens with empty input? - What happens with null / None / undefined values? - What happens at the boundaries (0, max int, empty string, empty list)? - What happens with unexpected data types? **Production assumptions** - What does this code assume about the environment that might not be true? - Does it assume network availability? File permissions? Specific OS? - Does it assume the API always responds correctly? **Security gaps** - Are there injection risks (SQL, command, path)? - Are credentials or secrets handled correctly? - Is user input ever trusted without validation? **Efficiency problems** - Are there N+1 query patterns? - Are there