From be646aae330bed268f00d69001382dce444ae635 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kacper=20Miko=C5=82ajczak?= Date: Mon, 23 Feb 2026 09:02:44 +0100 Subject: [PATCH] Add reality check instruction to AI code reviewer Instruction tells the reviewer to re-read the code before posting each inline comment and confirm the violation is real. If the code is actually correct, the comment is silently skipped instead of posted. Addresses ~23% of reviewed PRs receiving self-contradicting comments where the reviewer flags a violation then concludes it's not a problem. --- .claude/agents/code-inline-reviewer.md | 1 + 1 file changed, 1 insertion(+) diff --git a/.claude/agents/code-inline-reviewer.md b/.claude/agents/code-inline-reviewer.md index bddb0fd886ddd..05334cef22da8 100644 --- a/.claude/agents/code-inline-reviewer.md +++ b/.claude/agents/code-inline-reviewer.md @@ -61,6 +61,7 @@ Each rule file contains: 12. **DO NOT describe what you are doing, create comments with a summary, explanations, extra content, comments on rules that are NOT violated or ANYTHING ELSE.** Only inline comments regarding rules violations are allowed. If no violations are found, add a reaction instead of creating any comment. EXCEPTION: If you believe something MIGHT be a Rule violation but are uncertain, err on the side of creating an inline comment with your concern rather than skipping it. +13. **Reality check before posting**: Before creating each inline comment, re-read the specific code one more time and confirm the violation is real. If upon re-reading you realize the code is actually correct, **do NOT post the comment** — silently skip it and move on. Never post a comment that flags a violation and then concludes it is not actually a problem. ## Tool Usage Example