AgentSkillsCN

ai-critic

从配对评审中获取AI生成的审查建议，并据此修改代码以解决问题。当用户说“address AI feedback”、“address AI suggestions”、“fix AI review feedback”，或希望基于配对评审的AI分析结果对代码进行迭代优化时，即可调用此技能。

SKILL.md
--- frontmatter
name: ai-critic
description: >
  Fetch AI-generated review suggestions from pair-review and make code changes to address them.
  Use when the user says "address AI feedback", "address AI suggestions", "fix AI review feedback",
  or wants to iterate on code based on AI analysis results from pair-review.

AI Critic

Fetch AI-generated suggestions from pair-review and make code changes to address the valid ones.

Determine review context

Determine whether this is a local review or a PR review:

  1. •If the user explicitly says "local", use local mode.
  2. •Otherwise, determine the GitHub owner, repo, and PR number for the current branch. If a PR exists, use PR mode with repo and prNumber params.
  3. •If no PR exists, use local mode with path (absolute cwd) and headSha (git rev-parse HEAD) params.

Fetch AI suggestions

Call mcp__pair-review__get_ai_suggestions with the review context params. This returns suggestions from the latest analysis run by default.

If the user wants suggestions from a specific analysis run, call mcp__pair-review__get_ai_analysis_runs first to list available runs, then pass the appropriate runId to get_ai_suggestions.

Only active and adopted suggestions are included (dismissed ones are excluded).

If no suggestions are returned, tell the user there are no AI suggestions to address.

Triage and address suggestions

AI suggestions are not human-curated — apply judgment. For each suggestion:

  1. •Read the file at the referenced path and lines.
  2. •Evaluate the suggestion: is it a real issue, a false positive, or a stylistic preference?
  3. •If the suggestion is valid and actionable, make the code change.
  4. •If the suggestion is a false positive or not worth addressing, skip it and note why.

Use the ai_confidence field as a signal but not a hard threshold — low-confidence suggestions can still be valid.

Report

After processing all suggestions, provide a summary:

  • •How many suggestions were reviewed
  • •Which ones were addressed (and what changed)
  • •Which ones were skipped (and why)