Policy Diff Analyst

Compare two (or N) versions of a Chinese policy document, produce:

1. output/report.docx — analysis article in one of 3 styles (research / media / retail)
2. output/data.xlsx — 4-sheet data table (indicators / term frequency / strength scores / diff list)
3. output/charts/ — G1–G6 PNG charts

When this skill triggers

Trigger on any of:

• User provides two versions of a Chinese policy document and asks for comparison
• User says "对比最新/近 N 年的 XX 政策文件" (Mode B auto-fetch)
• User mentions 政府工作报告 / 中央经济工作会议 / 五年规划 / 三中全会决定 / 货币政策执行报告 comparison
• User wants policy-driven investment analysis (风格判断 / 行业配置 / 交易节奏)

Supported file_types

file_type	文档类型	profile 文件
govt_work_report	政府工作报告	references/profile_govt_work_report.md
cewc	中央经济工作会议	references/profile_cewc.md
five_year_plan	五年规划纲要	references/profile_five_year_plan.md
third_plenum	三中全会决定	references/profile_third_plenum.md
monetary_policy_report	货币政策执行报告	references/profile_monetary_policy_report.md

Required inputs (identify from the user's prompt; ask if missing)

Field	Mode A	Mode B	Default
new version	✅	—	—
old version	✅	—	—
intent description	—	✅	—
file_type	✅	✅	必须指定
N versions (Mode B only)	—	2	2
article_style	✅	✅	ask user; options: research / media / retail
output_dir	❌	❌	./output

Do not assume. If any required field is missing, ask one focused question.

Mode decision

• Mode A (explicit inputs): user supplies URLs, PDF/docx paths, or pasted text.
• Mode B (intent-driven): user describes intent like "最新两版" / "近三年趋势". You:
  1. Determine file_type, N, year range
  2. For each year, call scripts/source_registry.get_sources(file_type, year) to get template URLs
  3. Try template URLs first. If they 404, use source_registry.get_search_hint(file_type, year) with WebSearch to discover actual URLs
  4. Validate all discovered URLs with source_registry.is_allowed(url) — reject any URL outside the whitelist
  5. Display the URL list to the user and wait for confirmation before fetching
  6. After confirmation, fetch and parse. Only whitelisted URLs are ever fetched

Workflow

Step 0 — Structure outline & framework diff (auto-generated)

run_analysis automatically extracts the structure outline of both documents and compares them:

• Multi-level heading detection: L1 一、 / L2 （一） / L3 1.
• L1 + L2 two-level comparison: L1 checks top-level section changes; L2 checks subsection reordering within matched L1 pairs (e.g. "消费" promoted from （三） to （一） under "工作任务")
• Section ordering comparison: detects moved_up / moved_down / added / removed / renamed
• Framework-level summary (e.g. "L1 框架基本不变；L2 子章节：消费从第3升至第1")

This appears as the first section in annotated_new.docx ("一、框架结构对比") and in the structure_diff key of the result JSON. Present this to the reader before keyword-level analysis — it gives the most direct framework-level orientation.

Step 1 — Parse intent, collect inputs

Ask missing questions. For Mode B, present URL list and get confirmation.

Step 2 — Build config and invoke run.py

{
  "old": {"mode": "text|url|pdf|docx", ...},
  "new": {"mode": "text|url|pdf|docx", ...},
  "file_type": "govt_work_report",
  "profile_path": "references/profile_govt_work_report.md",
  "output_dir": "./output"
}

Save to /tmp/config.json, then run:

python -m scripts.run /tmp/config.json

This produces output/data.xlsx, output/charts/G1,G2,G6.png, and output/analysis_brief.json.

Alternatively, if installed via pip: pda compare --config /tmp/config.json

Step 3 — Read the analysis_brief.json

Open output/analysis_brief.json — this is the structured, LLM-friendly summary containing:

• structure_diff: L1/L2 framework changes (priority signal)
• top_changes: ranked keyword-level changes across all profile layers
• strength_scores: per-dimension policy strength with delta
• quantitative_indicators: numerical indicator changes
• new_term_candidates: high-frequency terms not in profile (for Step 11)

The full raw data is also available in run_analysis() return dict if needed.

Step 4 — Write the article body (YOU, the LLM)

First, load the analysis methodology:

• references/analysis_framework.md — 完整解读方法论（12 章）

Key framework sections to internalize:

• §一-§三：定调谱系 + 量化三步法 + 措辞四类信号
• §七：文档类型差异化视角——不同 file_type 的分析切入角度完全不同（如货政报告关注流动性措辞，政府工作报告关注量化目标）
• §八：关键词→行业传导映射表——政策关键词到 A 股板块的具体映射
• §九：政策周期定位——先判断当前处于稳增长/调结构/防风险/促发展哪个阶段
• §十：弦外之音读取——信心水平、紧迫度、政策语言翻译表
• §十一：浅层 vs 深度分析实例——必须达到"深度分析"示例的水平
• §十二：预期差分析——实际政策 vs 市场预期的偏差

Then, load the style guide:

• assets/style_research.md for research
• assets/style_media.md for media
• assets/style_retail.md for retail

Analysis requirements (all styles must cover, depth varies by style):

1. 政策周期定位：先判断当前处于什么周期阶段（§九），这是整篇分析的底色
2. 框架概览：从 structure_diff 提取 L1/L2 位置变化，呈现全局结构变动
3. 定调判断：按 §一 的谱系，判断宏观/财政/货币三重定调组合，与上年纵向对比
4. 量化目标解读：按 §二 的三步法（数字→设定逻辑→市场含义），解读核心指标变化
5. TOP 10 核心差异深度解读：每个按 §六.1 + §十一 的深度示例标准（差异项→信号解读→历史对标→传导推演→置信度），参考 §十 读取弦外之音
6. 行业传导：用 §八 的映射表，从政策差异推演到具体板块
7. C1 市场风格判断（§四.3）+ C2 行业配置（§四.2 + §八）+ C3 交易节奏
8. 预期差分析（§十二）：标注市场事前预期 vs 实际政策的偏差
9. 风险与反向观点：对核心判断给出反面论证 + §十.4 的反面验证

核心原则：每个差异项必须走完"是什么→为什么→所以呢"三步，裸列差异不是分析。参照 §十一 的深度分析示例——达不到该水平的分析不合格。

Hand the result to build_docx.build_report as sections: list[(heading, [paragraph_str, ...])]. The last section should be the appendix containing old and new document full text.

Step 5 — Produce G3 and G5 charts (LLM-derived data)

G3 config matrix and G5 sector sunburst depend on your C2 judgment. After writing the article:

• Build config dict for G3 from your C2 allocation
• Build sectors list for G5 from A3 industry hits
• Call build_charts.build_g3_config_matrix and build_charts.build_g5_sector_sunburst

Step 6 — Assemble report.docx

from scripts.build_docx import build_report
build_report(
    title="...",
    style="research",
    sections=[...],
    disclaimer="本文基于公开政策文件分析，仅供研究参考，不构成投资建议。",
    output_path="output/report.docx",
)

Step 7 — Annotated docx (auto-generated)

run_analysis automatically produces output/annotated_new.docx:

• Keywords added in new version: red bold + grey 【新增·A3】 inline tag
• Keywords with strength change: red bold + grey 【强度 3→5】 inline tag
• Keywords removed: collected in a "本版不再提及的表述" section at end

No extra config needed — this is always produced alongside data.xlsx.

Step 8 — Temporal diff (optional, N-year comparison)

If the user provides 3+ years of documents, add extra_docs to config:

{
  "old": {"mode": "text", "content": "...", "year": 2022},
  "new": {"mode": "text", "content": "...", "year": 2024},
  "extra_docs": [
    {"mode": "text", "content": "...", "year": 2023}
  ],
  ...
}

This produces temporal_report in the result dict with:

• Per-keyword trajectory: sustained / emerging / revived / fading / dropped
• strength_by_year: keyword strength scores across all years
• pairwise_reports: consecutive year-pair DiffReports

Use the temporal data to narrate multi-year trends in the article (e.g. "新质生产力 emerged in 2024, absent in 2022-2023").

Step 9 — Quantitative indicator comparison

run_analysis returns quantitative_comparison — a list of extracted numerical indicators (GDP targets, deficit ratios, bond issuance amounts, etc.) with old vs new values. Present these as a structured table in the article. Key signals: target changes (5%→5%), scale changes (3.9→4.4万亿), and newly introduced targets.

Step 10 — Report back

Tell the user:

• Output directory path
• What files were produced (including annotated_new.docx with structure outline)
• Structure framework changes first (from structure_diff.summary, both L1 and L2)
• Quantitative indicator changes (key number shifts)
• Then one-paragraph summary of keyword-level core findings

Step 11 — Profile self-expansion (auto-triggered)

After analysis, run_analysis returns new_term_candidates — a list of high-frequency terms found in the new document that are NOT in the current profile. Present these to the user:

"以下高频词不在当前 profile 关键词库中，是否需要加入？" (table of term / new_freq / old_freq)

If the user confirms specific terms, append them to the appropriate layer in the profile .md file. This keeps the keyword library evolving with policy vocabulary.

Step 12 — Self-evaluation (optional, recommended for quality iteration)

After completing the analysis article, perform a quality self-check:

Layer 2 — Expert cross-reference:

1. Use WebSearch: "{year}年政府工作报告 解读 中金/中信/国君" (or equivalent for other file_types)
2. Compare your core conclusions with expert commentary
3. Flag any signals experts caught that you missed, or signals you reported that experts didn't mention
4. Output a brief "分析质量自评" noting gaps and confidence level

Layer 3 — Temporal self-correction (for N-year analyses):

1. Check if previous year's temporal_report predictions held:
   • "emerging" keywords: did they persist? → precision check
   • "fading" keywords: did they actually drop? → recall check
2. Output prediction accuracy rate
3. Feed back into trajectory classification tuning

Constraints

• Never fetch URLs outside source_registry.ALLOWED_DOMAINS. The whitelist is hardcoded for a reason.
• Always ask user confirmation of Mode B URL list before fetching.
• Always include the disclaimer in the final report.
• Mode B max N = 5 versions.
• Article body: strictly follow the selected style guide — do not mix styles.

Cross-platform usage

The analysis engine is platform-agnostic. Three layers:

Layer	Content	Platform dependency
Engine (Python CLI)	scripts/ + pda compare CLI	None — any shell
Prompt template	prompts/analysis_prompt.md	None — any LLM
Platform adapter	SKILL.md / adapters/codex/AGENTS.md	Platform-specific

To use on a different AI platform:

1. Install: pip install -e .
2. Run: pda compare --old old.txt --new new.txt --file-type govt_work_report
3. Feed output/analysis_brief.json + prompts/analysis_prompt.md to your LLM
4. Pass the LLM output to scripts/build_docx.build_report() for Word generation

See adapters/codex/AGENTS.md for Codex-specific instructions.

Reference files

• prompts/analysis_prompt.md — universal analysis instructions (platform-agnostic)
• references/profile_govt_work_report.md — 政府工作报告 layer keyword map + scoring anchors
• references/profile_cewc.md — 中央经济工作会议 profile
• references/profile_five_year_plan.md — 五年规划纲要 profile
• references/profile_third_plenum.md — 三中全会决定 profile
• references/profile_monetary_policy_report.md — 货币政策执行报告 profile
• references/source_whitelist.md — human-readable domain whitelist
• assets/style_*.md — style guides (loaded per-invocation)
• adapters/codex/AGENTS.md — Codex platform adapter