Type: url
Original: https://note.com/hacklog_stealth/n/n395c7ca5da82
Status: generated
Language: ja
Summary
A note post by Hack-Log describing how the author stopped relying on Claude Code reviewing its own code and instead handed only the review step to a different model, Codex. It argues that same-model self-review inherits the implementer’s unstated assumptions, so an “adversarial review” prompt changes the reviewer’s posture but not its blind spots, and reports that a separate reviewer model surfaced assumption-level bugs such as a crash on a missing config file, swallowed errors and partially failed async work reported as success.