prd-per-analysis

name	prd-per-analysis
description	Analyze individual PRD documents across 6 dimensions with vision support for images. Use when (1) analyzing a PRD document's core problem, boundaries, and boundary-external issues, (2) tracking whether issues are resolved in later versions, (3) extracting key parameters and solution highlights from technical PRDs, (4) scoring document quality, (5) understanding flowcharts, architecture diagrams, UI mockups, or screenshots embedded in PRD documents. Also use when a user mentions 'per-doc analysis', 'boundary analysis', 'issue tracking', 'document review', or 'PRD quality scoring'.
license	MIT
compatibility	opencode

PRD Per-Document Analysis

Analyze each PRD document across 6 dimensions with optional vision (multimodal) support for embedded images.

This is the highest-value step in the review pipeline — "boundary-external issue tracking" reveals the evolution logic between requirements.

Dual Engine Architecture

This skill uses two LLM engines:

Engine	Purpose	Default Model	When Used
text_engine	Text analysis (6 dimensions, resolution tracking)	claude-sonnet-4-20250514	Always
vision_engine	Image understanding (diagrams, UI, flowcharts)	claude-sonnet-4-20250514	When --enable-vision and images exist

Override via environment variables:

• TEXT_MODEL / VISION_MODEL — override default models
• ANTHROPIC_API_KEY — required for both engines

Vision is opt-in because images consume significantly more tokens. When enabled, images are classified before sending to the vision engine to avoid wasting tokens on decorative images.

Image Classification

Image Type	Action	Examples
Flowchart / Architecture diagram	Send to vision engine	Decision trees, system architecture, sequence diagrams
UI / App screenshot	Send to vision engine	App pages, web interfaces, device screens
Data chart / Table image	Send to vision engine	Bar charts, pie charts, comparison tables
Decorative / Emoji / Icon	Skip	Decorative dividers, emoji, brand logos
Photo / Illustration	Context-dependent	User research photos may be relevant

Six-Dimension Analysis Framework

Dim	Name	What It Captures	Output Field
1	Core Problem	The specific problem this requirement solves	core_problem
2	Category	Which functional domain it belongs to	category
3	Boundary	What it does and doesn't cover	boundary_in[], boundary_out[]
4	Boundary-External Issues	Related problems NOT covered by this requirement	boundary_issues[]
5	Resolution Tracking	Whether boundary issues are resolved in later versions	boundary_issues[].resolution
6	Key Points Extraction	Type-specific highlights (technical/survey/competitive)	key_points

Dimension 6 type-specific rules:

Doc Type	What to Extract
Technical	Solution highlights + key parameters (thresholds, timeouts, algorithm names)
Survey	Research method + core insights
Competitive	Comparison dimensions + gap analysis

Quality Scoring

Score	Criteria
5	Clear boundaries, boundary-external issues acknowledged, parameters complete
4	Relatively clear boundaries, some boundary-external issues recognized
3	Basic boundaries exist, but insufficient boundary-external issue identification
2	Vague boundaries, missing key parameters
1	No boundary definition, core problem unclear

Quick Start

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

Analyze a single document

python scripts/analyze.py <md_path> <output_json> [options]

# Basic text-only analysis
python scripts/analyze.py /path/to/doc.md result.json

# With category and version from prd-overview-classify
python scripts/analyze.py /path/to/doc.md result.json --category "核心策略" --version "V2.3.6"

# Enable vision for images in assets/ folder
python scripts/analyze.py /path/to/doc.md result.json --enable-vision

# With other docs' excerpts for resolution tracking
python scripts/analyze.py /path/to/doc.md result.json --context context.json

Analyze a batch of documents

python scripts/batch_analyze.py <classify_result_json> <output_dir> [options]

# Basic batch analysis (text only)
python scripts/batch_analyze.py classify-result.json ./analysis-results/

# With vision enabled
python scripts/batch_analyze.py classify-result.json ./analysis-results/ --enable-vision

# Control concurrency (default: 3)
python scripts/batch_analyze.py classify-result.json ./analysis-results/ --max-concurrent 5

Input

Single document mode

The analyze.py script takes a Markdown file path and optional context.

Batch mode

The batch_analyze.py script takes the output JSON from prd-overview-classify as input, which contains document paths, categories, and version chains needed for resolution tracking.

Context JSON (for resolution tracking)

{
  "other_docs_excerpts": [
    {
      "doc_id": "abc123",
      "version": "V2.3.5",
      "title": "智能判定下发策略V2",
      "boundary_issues": ["edge case not handled"]
    }
  ]
}

Output Format

{
  "doc_id": "abc123",
  "core_problem": "新旧算法设备混合组网的判定流程",
  "category": "核心策略",
  "boundary_in": [
    "新旧算法混合判定流程",
    "算法版本标记机制"
  ],
  "boundary_out": [
    "纯新/纯旧算法场景（已有独立流程）"
  ],
  "boundary_issues": [
    {
      "issue": "混合组网时边缘情况的权重配比未详细定义",
      "severity": "medium",
      "resolution": {
        "status": "unresolved",
        "resolved_by": null,
        "evidence": null,
        "note": "当前版本未覆盖此场景"
      }
    }
  ],
  "key_points": {
    "type": "technical",
    "solution_highlights": ["NewTop vs OldTop×0.7 对比", "3阶段过渡方案"],
    "key_parameters": [
      {"name": "混合策略阈值", "value": "0.7"},
      {"name": "A/B测试周期", "value": "2周"}
    ]
  },
  "image_insights": [
    {
      "image_path": "assets/image3.png",
      "image_type": "flowchart",
      "description": "新旧算法判定的3阶段流程：纯新→纯旧→混合",
      "relevant_dimensions": ["core_problem", "boundary_in"]
    }
  ],
  "quality_score": 4.0,
  "confidence": 0.92
}

Key Features

• Dual engine: Text engine for analysis, vision engine for images (opt-in)
• Image classification: Automatically filters decorative images before sending to vision engine
• Resolution tracking: Cross-references boundary issues with other documents
• Type-aware extraction: Different extraction rules for technical/survey/competitive docs
• Batch processing: Concurrent analysis with configurable rate limiting
• Incremental: Analyze only new or changed documents

Dependencies

pip install -r requirements.txt
# Core: anthropic, pydantic
# Optional: Pillow (image format detection)

Scripts Reference

Script	Purpose
analyze.py	Single document analysis (text + vision)
batch_analyze.py	Batch analysis with concurrency control

Prompts Reference

Prompt	Purpose
per-doc-analysis.md	6-dimension analysis (text + vision)
resolution-tracking.md	Boundary issue resolution tracking

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