AgentSkillsCN

review-docs

通过为每份文档指派文档策展代理,对文档进行细致的审核,确保其与代码库完全匹配。

SKILL.md
--- frontmatter
name: review-docs
description:
  Orchestrate documentation review by assigning docs-curator agents to
  individual documents for thorough validation against the codebase.

Documentation Review Orchestration

A management skill for systematically reviewing documentation. Assign one docs-curator agent per document to ensure thorough investigation.

When to Use

Activate when:

  • User wants to audit documentation for accuracy
  • User asks to check if proposals/plans have been implemented
  • User wants to identify outdated documentation
  • User mentions "review docs", "check proposals", or "documentation audit"
  • After major features are completed and docs need validation

Key indicators:

  • "Review the proposals folder"
  • "Check which plans are complete"
  • "Is the architecture documentation still accurate?"
  • "Audit the docs for outdated content"

Core Principle: One Agent Per Document

Assigning multiple documents to a single agent dilutes review quality. Each document requires:

  • Full document reading
  • Codebase searching for evidence
  • Git history investigation
  • Cross-referencing related docs

Always assign one docs-curator agent per document.

Workflow

Step 1: Identify Documents to Review

List the documents that need review:

bash
# List project proposals
ls docs/projects/*/proposal.md

# List project plans
ls docs/projects/*/plan.md

# List architecture docs
ls docs/architecture/*.md

# List specifications
ls docs/specifications/*.md

Or use Glob to find specific patterns:

  • docs/projects/*/proposal.md - All project proposals
  • docs/projects/*/plan.md - All project plans
  • docs/projects/*/sessions/*.md - All project sessions
  • docs/architecture/*.md - All architecture docs
  • docs/specifications/*.md - All specifications

Step 2: Categorize by Document Type

Different document types have different review goals:

Temporal Documents (check for completion/archival):

  • docs/projects/*/proposal.md - Archive project when implemented
  • docs/projects/*/plan.md - Archive project when complete
  • docs/investigations/ - Archive when concluded
  • docs/reports/ - Archive when acted upon

Evergreen Documents (check for accuracy):

  • docs/architecture/ - Update to match current code
  • docs/specifications/ - Update to match current application behavior
  • docs/playbooks/ - Update if procedures changed
  • docs/lessons-learned/ - Update if solutions evolved
  • docs/interaction-design/ - Update to match current UX

Step 3: Assign Agents

Launch one docs-curator agent per document using the Task tool:

code
Task(
  subagent_type: "docs-curator",
  description: "Review [document-name]",
  prompt: "Review this document for accuracy and implementation status:

  **Document:** docs/projects/feature-x/proposal.md

  Read the document, search the codebase for evidence of implementation,
  check git history for related commits, and provide your findings with
  the recommended action (archive, update, no action)."
)

For parallel efficiency, launch multiple agents in a single message:

code
// Launch 4 agents in parallel, one per document
Task(...doc1...)
Task(...doc2...)
Task(...doc3...)
Task(...doc4...)

Step 4: Collect and Consolidate Results

As agents complete, collect their findings:

  1. Group by recommended action:

    • Ready to archive (100% complete)
    • Needs status update (partially complete)
    • Needs content update (outdated sections)
    • No action needed (current and accurate)
  2. Present summary to user:

    markdown
    ## Documentation Review Results
    
    ### Ready to Archive (5 documents)
    
    - projects/feature-x/ - Fully implemented
    - projects/feature-y/ - Superseded by feature-z ...
    
    ### Needs Update (2 documents)
    
    - architecture/api-design.md - Section 3 outdated ...
    
    ### Current (3 documents)
    
    - playbooks/deployment.md - Accurate ...
    
  3. Get user approval before taking action

Step 5: Execute Approved Actions

For approved archives:

  1. Update status in document
  2. Move to archive/ subfolder

For approved updates:

  1. Make the specific changes identified
  2. Or create tasks for the user to address

Batch Size Guidelines

For large documentation sets:

Doc CountApproach
1-5Review all in parallel
6-15Batch into 2-3 waves
16+Prioritize by age/importance, review subset

Prioritization criteria:

  • Oldest documents first (most likely outdated)
  • Documents related to recently completed features
  • Documents the user specifically mentioned

Agent Assignment Template

Copy and customize for each document:

code
Review this document for accuracy and implementation status:

**Document:** [full path to document]
**Document Type:** [Proposal | Plan | Architecture | etc.]

Tasks:
1. Read the full document
2. Search codebase for mentioned features/components
3. Check git log for related commits
4. Determine implementation/accuracy status
5. Recommend action: Archive | Update | No Action

Provide findings in the standard docs-curator output format.

Example: Reviewing All Project Proposals

bash
# 1. List active project proposals (excluding archive)
ls docs/projects/*/proposal.md | grep -v archive

# Result: 8 projects to review
code
# 2. Launch 4 agents in parallel (first batch)
Task(subagent_type: "docs-curator", description: "Review project-a proposal", prompt: "...")
Task(subagent_type: "docs-curator", description: "Review project-b proposal", prompt: "...")
Task(subagent_type: "docs-curator", description: "Review project-c proposal", prompt: "...")
Task(subagent_type: "docs-curator", description: "Review project-d proposal", prompt: "...")
code
# 3. After first batch completes, launch remaining 4
Task(subagent_type: "docs-curator", description: "Review project-e proposal", prompt: "...")
...
code
# 4. Consolidate results and present to user
"8 projects reviewed:
- 5 ready for archive (fully implemented)
- 2 partially complete (list what remains)
- 1 not started

Would you like me to archive the 5 completed project folders?"

Checklist: Documentation Review

  • Identify documents to review (list files)
  • Categorize by document type
  • Launch one docs-curator agent per document
  • Wait for all agents to complete
  • Consolidate findings by recommended action
  • Present summary to user
  • Get approval before archiving/updating
  • Execute approved actions
  • Report completion