Custom Devin review skill reportedly caught about 70% of a developer’s flagged code issues in a held-out test

A developer says they built a Devin skill to review pull requests in their style, refining it with comments from six months of GitHub PRs and Devin sessions. In a held-out evaluation, they say it caught about 70% of the code issues they had flagged. They reserve product decisions for themselves and say the skill sorts issues into auto-fixes and those needing their approval. The overnight refinement used 91 subagents, 559 grandsubagents and 1,883 great-grandsubagents.