AgentSkillsCN

plan-review

当用户说“审查”时,进行全面的方案审核。扫描方案,查找漏洞、改进点、创新之处以及潜在缺口。当用户输入“审查”、“审查方案”或要求进行全方位、高精度的方案审计时使用。

SKILL.md
--- frontmatter
name: plan-review
description: Perform exhaustive plan review when user says "review". Scans plans for flaws, improvements, innovations, and gaps. Use when user types "review", "review the plan", or asks for a comprehensive plan audit with maximum scrutiny.

Plan Review (Maximum Scrutiny Mode)

When the user invokes the review keyword (e.g., "review", "review the plan", "run a review"), enter the highest level of alertness and perform a comprehensive plan audit.

Trigger

  • User says "review" or "review the plan" in context of a plan document
  • User asks for a thorough audit, scrutiny, or critical analysis of a plan

Review Protocol

1. Full Read

  • Read the entire plan file from start to finish
  • Load any referenced files (configs, scripts, docs) that the plan depends on
  • Re-read sections that reference external systems, APIs, or integrations

2. Flaw Detection

Scan for:

  • Logical gaps: Missing steps, circular dependencies, undefined terms
  • Contradictions: Conflicting statements, incompatible assumptions
  • Underspecified links: Handoffs between components that lack clear interfaces (e.g., "calls OpenClaw" without specifying how output is captured)
  • Unstated assumptions: Requirements implied but not written
  • Overlooked edge cases: Failure modes, timeout scenarios, partial success
  • Platform/OS blind spots: Windows vs Unix paths, shell differences, env differences

3. Improvement Opportunities

Identify:

  • Clarity: Ambiguous phrasing, unclear ownership, vague success criteria
  • Completeness: Missing phases, skipped prerequisites, absent rollback/undo
  • Feasibility: Overly optimistic timelines, underestimated complexity, missing dependencies
  • Maintainability: Hardcoded values, lack of versioning, missing logging/observability
  • Security: Credentials, secrets, PII handling

4. Innovation & Optimization

Consider:

  • Alternative approaches: Simpler architectures, different tool choices
  • Automation opportunities: Steps that could be further automated
  • Cost or time optimizations: Cheaper or faster paths to the same outcome
  • Future-proofing: Extensibility, modularity for later phases
  • Industry best practices: Patterns the plan could adopt

5. Situational Context

  • Project type: Is this a greenfield build, migration, integration, or refactor?
  • Stakeholders: Who depends on this? What happens if it slips?
  • Environment: Local, cloud, hybrid? Any constraints (budget, compliance, legacy)?
  • Dependencies: External APIs, third-party services, rate limits, SLAs

6. Handle Findings

  • Report all findings in a structured format (Flaws, Improvements, Innovations)
  • Prioritize: Critical (blocks success) vs High (should fix) vs Medium (nice to have) vs Low (optional)
  • Propose fixes: For each finding, suggest a concrete change to the plan
  • Apply changes: Update the plan document with approved fixes unless the user prefers review-only

Output Format

markdown
## Plan Review Summary

**Plan:** [name/path]
**Reviewed:** [date]

### Critical Flaws
- [Finding] → [Proposed fix]

### Improvements
- [Finding] → [Proposed fix]

### Innovations
- [Suggestion] → [Rationale]

### Applied Changes
- [List of edits made to the plan]

Intensity Level

Treat "review" as a signal to maximize cognitive effort. Do not shortcut. Question every link in the chain. Assume the plan will be executed by someone who has not read the implicit context.