charted-review

📁 marmicode/skills 📅 12 days ago
9
总安装量
9
周安装量
#31619
全站排名
安装命令
npx skills add https://github.com/marmicode/skills --skill charted-review

Agent 安装分布

cursor 9
gemini-cli 8
github-copilot 8
opencode 7
codex 7
amp 7

Skill 文档

Context

  • designDocPath: $ARGUMENTS[0]

Goal

Review the design doc at ${designDocPath} by dispatching four expert sub-agents in parallel, then synthesize their feedback into a single, actionable review.

Step 1 – Set up agents

If the sub-agents from Step 2 are not already configured, ask me to install them. If I confirm, copy sub-agents from the ./assets/agents folder to the {workspaceRoot}/.cursor/agents folder if using Cursor. Adapt the paths if I am not using Cursor.

Step 2 — Parallel Expert Reviews

Spawn the following four sub-agents simultaneously (all in one message). Each agent receives the full design doc content and must return a structured review.

Sub-agent type Focus area
accessibility-expert Accessibility of the proposed UI — ARIA, keyboard navigation, screen-reader support, color contrast, focus management, semantic HTML.
security-analyst Security implications — input validation, injection risks, auth/authz gaps, data exposure, secure defaults.
ux-expert Usability — interaction patterns, error/empty/loading states, information architecture, responsiveness, cognitive load.
xp-coach Engineering practices — testability, incremental delivery, PR plan quality, simplicity, YAGNI, refactoring opportunities.

Each sub-agent prompt must include:

  1. The full content of the design doc.
  2. Instructions to return a review with exactly these sections:
    • Praise — what the design does well (keep brief).
    • Concerns — numbered list. Each concern has a short title, an explanation, and a concrete suggestion.
    • Verdict — one of: approve, request-changes, or needs-discussion.

Step 3 — Synthesize & Detect Conflicts

Collect the four reviews. Present a summary table to the user:

Expert Verdict # Concerns

Then list all concerns grouped by expert.

After listing, identify conflicts — cases where two or more experts give contradictory guidance (e.g., one says “add a confirmation dialog” and another says “reduce interaction steps”, or one says “split into more PRs” and another says “too many PRs already”).

Step 4 — Challenge Round (if conflicts exist)

For each conflict:

  1. Clearly describe the disagreement to both involved experts.
  2. Re-launch each conflicting sub-agent with:
    • The original design doc.
    • Their own original review.
    • The opposing expert’s concern that contradicts theirs.
    • Instructions to either revise their position or defend it with stronger justification.
  3. Run conflicting pairs in parallel when they are independent.

After the challenge round, check if the experts now agree.

Step 5 — Resolution

  • If all conflicts are resolved: present the final consolidated review with the agreed-upon changes.
  • If any conflict remains unresolved: present the remaining disagreement(s) to the user in a clear format:

Unresolved: {short title}

{Expert A} argues: {summary of position}

{Expert B} argues: {summary of position}

What is your call?

Wait for the user’s decision on each unresolved conflict before producing the final consolidated review.

Step 6 — Final Output

Produce a consolidated review with:

  1. Overall Verdict — approve, request-changes, or approve-with-nits based on the aggregated outcome.
  2. Action Items — a numbered checklist of concrete changes to make to the design doc, ordered by priority.
  3. Resolved Conflicts — brief note on how each conflict was settled (expert concession or user decision).