PR Review Address Feedback Skill

Overview

This skill processes PR feedback by implementing a Validate-Then-Act pattern: fetch all comments from every GitHub source, independently verify each claim, present findings in a summary table, then fix only validated issues. The pattern is mandatory because:

• Reviewer claims are hypotheses, not facts (verify with code/tests/HTTP requests)
• Single comment sources miss feedback (reviews, inline comments, and issue comments are separate endpoints)
• Unverified fixes waste time and introduce bugs (validation phase gates code changes)

Key Principles

Fetch Comprehensively: Always query all three endpoints — reviews, inline comments, AND issue comments. Claude and other tools comment via /issues/{pr}/comments, which is a separate endpoint. Missing any source means missing feedback.

Verify Independently: Reviewer claims must be tested using the verification tests table. Reviewers often miss indirect usage, re-exports, or dynamic references. Trust hierarchy (highest to lowest): running code/tests → HTTP requests → grep/search → reading source → reviewer's word.

Classify Honestly: Classify comments as VALID, INVALID, or NEEDS-DISCUSSION based on evidence. Never downgrade NEEDS-DISCUSSION to VALID to avoid asking the user — acting on ambiguous feedback without user input may fix the wrong thing.

Present Summary Before Fixes: Complete all validations and present the full summary table before making any code changes. User loses the big picture if you validate and fix incrementally, and NEEDS-DISCUSSION items may change approach to VALID ones.