Claude Code Plugins

Community-maintained marketplace

Feedback

review-diffs

@uuta/dotfiles
1
0

Performs comprehensive code review on git diffs and generates review.md with quality scores, critical issues, performance concerns, and potential bugs. Use when user requests to review changes, check code quality, analyze diffs, or validate LLM-generated code

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name review-diffs
description Performs comprehensive code review on git diffs and generates review.md with quality scores, critical issues, performance concerns, and potential bugs. Use when user requests to review changes, check code quality, analyze diffs, or validate LLM-generated code

Review diffs

Instructions

  • Read diffs with git diff commands, and review code after LLM changes code

Todo

  • Read diffs with git diff commands
  • Review diffs
  • Create/update docs/review.md in the current directory that you're working on

What docs/review.md should contain

  • Score (up to 100)
  • Short summary
  • Critical issue
  • Potentially performance issue
  • potential bug

Examples

  • Score (up to 100): 45
  • Short summary: format_all_attachment is now wired into the token counter, but the prompt builder still falls back to the old helper that drops full-file attachments whenever a ranged snippet exists, so the shipped behavior hasn’t actually improved and the token math is now inconsistent with what the LLM sees.
  • Critical issue:
    • PromptBuilder.build_prompt still chooses format_ranged_attachment for the whole batch whenever any attachment has line_range, and the helper only emits entries for those ranged files, so attachments without ranges vanish from the actual prompt even though the new formatter counts them. Users still lose files and won’t get their code reviewed. (src/usecase/message/prompt_builder.py:24, src/usecase/helper/helper.py:198)
  • Potentially performance issue:
    • GenerateMessageUseCase.execute now counts tokens for every attachment via format_all_attachment, but the prompt builder sends fewer tokens when the mix contains both ranged and full files. Legitimate requests are rejected too early with EEAOAI40003 even though the downstream prompt would have fit, effectively shrinking usable context and wasting retries. (src/usecase/message/generate.py:140, src/usecase/message/prompt_builder.py:24)
  • potential bug:
    • The same prompt-builder helper is used for each HistoryItem, so any historical message that combined ranged and non-ranged attachments will also drop the non-ranged ones. Past context can silently disappear between turns, leading the model to hallucinate or ignore previous uploads. (src/usecase/message/prompt_builder.py:33)