prd-overview-classify

name	prd-overview-classify
description	Classify and overview a batch of Markdown PRD documents by category, build version evolution chains, and detect inter-document dependencies. Use when categorizing PRD or requirement documents by topic, extracting version numbers, finding document relationships, or generating overview summaries. Also use when you need to understand the structure of a document set before deep analysis, or when a user mentions 'overview', 'categorize', 'classify documents', 'version chain', or 'document relationships'.
license	MIT
compatibility	opencode

PRD Overview & Classification

Automatically classify PRD documents by category, extract version numbers, build version chains, and identify inter-document dependencies.

Workflow Overview

1. Document scanning    → Extract metadata (version, title, excerpt)
2. Classification       → Assign each doc to a category (keyword + LLM)
3. Version chain build  → Link docs in same evolution line
4. Dependency detect    → Find cross-doc references/resolutions
5. Summary output       → Generate structured JSON result

Quick Start

All commands below assume the working directory is the skill root (skills/prd-overview-classify/). Install dependencies first: pip install -r requirements.txt

Classify a batch of Markdown documents

python scripts/classify.py <input_dir> <output_json> [options]

# Basic usage
python scripts/classify.py /path/to/converted_docs result.json

# With custom category config
python scripts/classify.py /path/to/converted_docs result.json --categories templates/default-categories.json

# With LLM fallback for uncertain classifications (requires ANTHROPIC_API_KEY)
python scripts/classify.py /path/to/converted_docs result.json --use-llm

# Override version pattern
python scripts/classify.py /path/to/converted_docs result.json --version-pattern "V\\d+\\.\\d+\\.\\d+"

# Skip LLM, use keyword-only classification (faster, no API cost)
python scripts/classify.py /path/to/converted_docs result.json --keyword-only

# Include document excerpts in output (disabled by default to keep JSON small)
python scripts/classify.py /path/to/converted_docs result.json --include-excerpts

# Customize excerpt length for LLM context
python scripts/classify.py /path/to/converted_docs result.json --excerpt-lines 300

Input directory structure

Expected input is the output of docx-to-markdown skill:

input_dir/
├── DocumentName1/
│   ├── DocumentName1.md
│   └── assets/
├── DocumentName2/
│   ├── DocumentName2.md
│   └── assets/
└── ...

Output Format

{
  "categories": [
    {
      "name": "核心策略",
      "doc_count": 11,
      "doc_ids": ["doc1", "doc2"]
    }
  ],
  "version_chains": [
    {
      "chain_name": "响应时延",
      "versions": [
        {"version": "V1.8.0", "doc_id": "doc1", "title": "智能联动响应时延V1"},
        {"version": "V2.1.0", "doc_id": "doc2", "title": "智能联动响应时延V2"}
      ]
    }
  ],
  "dependencies": [
    {
      "from_doc_id": "doc2",
      "to_doc_id": "doc1",
      "relation": "version_successor",
      "description": "V2.1.0 is successor to V1.8.0 in chain '响应时延'"
    }
  ],
  "documents": [
    {
      "doc_id": "doc1",
      "filename": "智能联动V2.3.6—智能判定流程V3",
      "md_path": "/path/to/doc.md",
      "category": "核心策略",
      "version": "V2.3.6",
      "subcategory_name": "核心策略",
      "subcategory_seq": 21,
      "title": "智能判定流程V3—兼容新旧算法",
      "line_count": 122,
      "file_size": 156000
    }
  ],
  "summary": {
    "total_docs": 29,
    "total_categories": 5,
    "total_chains": 5,
    "total_dependencies": 3
  }
}

Note: The excerpt field in documents is excluded by default to keep output JSON compact. Use --include-excerpts to include it. Downstream skills should read the original .md files directly via md_path.

Classification Strategy

Two-phase approach

1. Keyword matching (fast, deterministic): Match document titles against configurable keyword lists
2. LLM semantic classification (accurate, costs API calls): For documents not matched by keywords, use LLM to classify based on title + excerpt

When --keyword-only is used, unmatched documents get category "未分类".

Default category keywords

See templates/default-categories.json. For domain-specific examples (smart home, SaaS, etc.), see references/category-examples.md.

Custom categories

Create a JSON file with your own category definitions:

{
  "categories": [
    {
      "name": "Your Category",
      "keywords": ["keyword1", "keyword2"],
      "description": "What this category covers"
    }
  ],
  "version_pattern": "V\\d+\\.\\d+[\\.\\d]*",
  "subcategory_pattern": "【(.+?)v(\\d+)】"
}

Then pass it with --categories your-categories.json.

Version Chain Detection

Two strategies, used in order:

1. Subcategory tag clustering: Extract tags from brackets like 【核心策略v21】, group by same tag name
2. Name prefix clustering (fallback): When no subcategory tags, group by filename prefix before version number (e.g., "功能预约V2.0" and "功能预约V2.1" → same chain)

Steps:

1. Extract version numbers from filenames using regex (configurable)
2. Extract subcategory tags (if present)
3. Cluster documents by semantic topic
4. Sort by version number within each cluster
5. When --use-llm is enabled, LLM validates whether semantically related docs belong to the same chain

Key Features

• Configurable categories: Override default keywords via JSON config file
• Version pattern flexibility: Regex for version extraction is configurable
• Compact output: Excerpts excluded by default, included on demand
• No LLM required for basic use: Keyword-only mode works offline with zero API cost
• Dual chain strategy: Subcategory tags + name prefix fallback for different doc naming conventions

Dependencies

pip install -r requirements.txt
# Core: pydantic (data validation)
# Optional: anthropic (LLM fallback classification)

Set ANTHROPIC_API_KEY environment variable for LLM mode. Set ANTHROPIC_MODEL to override the default model (default: claude-sonnet-4-20250514).

Scripts Reference

Script	Purpose	Dependencies
classify.py	Main entry: scan + classify + chain build + output	requirements.txt

Prompts Reference

Prompt	Purpose	Used When
classify.md	LLM document classification	--use-llm and keyword unmatched
version-chain.md	LLM version chain validation	--use-llm and chain ambiguous

For detailed API and customization, see references/usage-guide.md.