Dairy Market Report

Generate a comprehensive, magazine-style PDF report summarizing monthly Chinese dairy market data from industry PDFs and policy documents in 12 structured se...

Dairy Market Report (乳制品市场行情报告)

Overview

This skill ingests one or more Chinese-language dairy industry monthly PDFs (e.g. 艾格农业《中国乳业研究月报》) and any related .docx policy briefs in the working directory, extracts the relevant data, and produces a complete, multi-section 乳制品市场行情报告 as a single PDF document (乳制品市场行情报告_YYYYMM.pdf).

The report follows a fixed 12-section structure (see references/report-template.md) and embeds GDT-to-warehouse landed-cost calculations per the user's formulas.

Visual identity (magazine-style, continuous flow)

The default PDF aesthetic is a professional business-magazine look — not a default reportlab dump. Concretely:

• Cover page (page 1): deep navy band with the report title + period in gold, followed by a米色 "CORE TAKEAWAYS · 核心结论" card with bulleted one-liners.
• Two-column TOC under the cover, listing all 12 sections.
• Body sections flow continuously — no PageBreak per section. The renderer (reportlab + STSong-Light CJK font) handles natural pagination. Typical output is 8–10 pages.
• Section header: SECTION 0X / 12 ordinal on the left, large dark-navy title, a short gold rule under it.
• KPI cards: 3–4 per section, each = small label + large value + colored change pill (e.g. +3.8% red ↑ or -9.4% green ↓).
• Data tables: navy header band, banded rows (white / 米 alternating), soft 0.25pt borders, numeric right-aligned.
• Change pills: filled with light red/green/grey, same-color border, ↑/↓/→ glyph.
• AI 解读 callout: 3pt gold left bar + 浅米 background, italic 解读 line.
• Source footer under every data-bearing section in small grey type.
• Page footer: hairline rule + report title + — N — page number on every body page.

Color tokens (use these exactly; renderer has them baked in):

Token	Hex	Use
INK	#1B2733	body text
NAVY	#0E2A47	primary brand, section titles, table header
ACCENT	#D4A24C	gold accents, callout bar, cover subtitle
SOFT_BG	#F4F1EA	callout background, zebra row B
RED	#B0322B	price-up (Chinese stock convention)
GREEN	#1F7A4D	price-down

Do not insert a hard PageBreak between sections, and do not wrap every section in KeepInFrame. The renderer already balances density; forcing a page break per section leaves large white gaps.

Data schema (what render_pdf.py consumes)

render_pdf.py --data <data.json> expects a single JSON object. The schema lives at the top of scripts/render_pdf.py and is mirrored in references/data-schema.md. The minimum fields are:

json

{
  "meta": {
    "title": "乳制品市场行情报告",
    "period": "2023年10月",
    "generated_at": "2026-06-26",
    "sources": ["艾格农业《中国乳业研究月报》202310", "农业农村部"]
  },
  "key_takeaways": ["国内奶价 ...", "GDT 反弹 ...", "..."],
  "sections": [
    { "title": "乳业整体形势概览", "kpis": [...], "tables": [...], "callout": "...", "source": "..." },
    { "title": "生鲜乳",          "kpis": [...], "tables": [...], "callout": "...", "source": "..." },
    ...
    { "title": "乳业重要事件及乳品企业动态资讯", "events": [...], "source": "..." }
  ]
}

Each kpis entry is {label, value, change} (change string is parsed for +/- or ↑/↓ to pick the pill color). Each tables entry is {header: [str], rows: [[str, ...]]}. callout is a single string.

When to use this skill

Trigger this skill when the user asks for any of the following, especially when Chinese-language dairy PDFs are present in the working directory:

• 生成乳制品市场行情报告 / 乳业月报 / 乳业研究月报
• 解读艾格农业《中国乳业研究月报》/ 制作对应行情报告
• 跑一遍乳业月报 → 输出 PDF
• 生鲜乳 / 饲料 / GDT / 进口乳制品 / 乳品深加工 综合行情

Do NOT use this skill for: ad-hoc single-data-point questions, English-language dairy market reports (use the English-language equivalent skill), or topics unrelated to dairy commodities.

Bundled resources

Read these on demand — they are not all needed up front.

File	Read it when
references/report-template.md	You are about to draft the report — it defines each of the 12 sections, expected data points, narrative guidance, and a "deep takeaway" expectation per section
references/data-schema.md	You are about to assemble the data JSON for the renderer — it documents every key (meta, key_takeaways, sections[*].kpis/tables/callout/source/events) and the change-pill sign convention
references/data-sources.md	You need to look up a number but the PDF didn't spell it out — covers 农业农村部, 国家统计局, 海关总署, GDT, USDA, Dairy Australia, Eurostat, DCANZ, AHDB, etc.
references/gdt-formulas.md	You are filling section 8 (GDT → 入仓成本) — restates the user's formulas and shows the example worked-out numbers
scripts/extract_pdf_text.py	You need raw text from a .pdf (e.g. 艾格农业月报). Usage: python extract_pdf_text.py <pdf> [out.txt]
scripts/extract_docx_text.py	You need raw text from a .docx (e.g. 政策文件). Usage: python extract_docx_text.py <docx> [out.txt]
scripts/gdt_calculator.py	You are computing section 8 numbers. Usage: python gdt_calculator.py --contract2-amf <price> --contract2-wmp <price> --contract2-smp <price> --fx <CNY/USD> --out <csv>
scripts/render_pdf.py	You have the filled data dict and need to emit the magazine-style PDF. Usage: python render_pdf.py --data <data.json> --out <report.pdf>. Uses reportlab + built-in CJK font (STSong-Light); no Chromium/Playwright required. Do not modify the layout logic in this file — it is the canonical visual spec for the report.
assets/styles.css	(DEPRECATED) — kept for reference only; the PDF is generated natively and does not need CSS
assets/report-template.html	(DEPRECATED) — kept for reference only; the PDF is generated natively and does not need an HTML template

Workflow

Follow these steps in order. Stop and ask the user only if a required input is genuinely missing — most inputs are recoverable from the PDFs themselves.

1. Discover input documents

List the working directory (ls / Glob for *.pdf, *.docx, *.txt, *.md, *.xlsx, *.csv). For each:

• PDF that looks like a dairy monthly (e.g. 艾格农业《中国乳业研究月报》YYYYMM) → primary source, must extract
• DOCX with policy / industry-event language → supplementary source
• Other files (CSVs, Excel) → may be raw data dumps; peek inside if small

Ask the user only if the report target month / year is not obvious from the filenames.

2. Extract text from all PDFs and DOCX

Run the bundled extractors on every relevant file:

bash

python scripts/extract_pdf_text.py "艾格农业-中国乳业研究月报202310.pdf" work/202310.txt
python scripts/extract_docx_text.py "农业农村部召开科技创新....docx" work/policy.txt

Place the text files in a work/ subfolder (scratch space). Read them fully (or in slices) to identify which sections of references/report-template.md each PDF covers. One monthly typically covers raw milk, feed, production, imports, exports, GDT, and policy; a single PDF rarely has everything, so cross-check.

3. Fill the 12-section report

Open references/report-template.md and references/data-sources.md. For each section:

1. Read the section spec (what data points are mandatory)
2. Find the matching numbers in the extracted text
3. If a number is missing in the PDF, consult data-sources.md for the official source URL — but do not fabricate. Mark it as "数据待补充" (data TBD) if truly unavailable.
4. Write the prose in Chinese, with English/Latin terms (GDT, AMF, WMP, SMP, YTD, YoY, MoM) allowed.
5. End each data-heavy section with a 1–3 sentence AI 解读 / 关键洞察 callout.

Special handling for section 8 (GDT → 入仓成本):

• Identify the most recent GDT event from the PDF (event number, date, contract-2 results for AMF, WMP, SMP).
• Use current CNY/USD midpoint from the PDF (or data-sources.md). Mark the FX source.
• Call scripts/gdt_calculator.py with those three contract-2 prices + FX to produce the landed-cost table.
• Embed the resulting CNY/吨 numbers in the report, with the formula in a footnote.

Special handling for section 11 (AI 深度解读 key takeaways):

• This section is purely synthesis. Read every other section first, then write 5–8 numbered insights that connect dots across sections (e.g. "国内生鲜乳价格 X 月连降 + GDT 拍卖连续 Y 次下跌 → 进口替代窗口期打开").
• Tie each insight to a number, not a vibe.

4. Build a data structure for the PDF

Assemble the filled content into a single JSON object matching the schema in references/data-schema.md (mirrored in the docstring of scripts/render_pdf.py). Top-level keys:

• meta — {title, period, generated_at, sources} (cover band + footer)
• key_takeaways — list of 3–6 short strings, bulleted on the cover card
• sections — ordered list of 12 entries, each with:
  • title (Chinese)
  • kpis — list of {label, value, change}; change is a free-form string like +3.8% or -9.4% YoY; the renderer auto-colors it (red for +/↑, green for -/↓, grey otherwise)
  • tables — list of {header: [str, ...], rows: [[str, ...], ...]}
  • callout — a single string, becomes the gold-bar 解读 box
  • source — a single string, becomes the grey footer line
  • (last section may also carry events — a list of {date, title, body} cards)

Rules:

• Currency: CNY 元 / 吨 as the default unit; show USD alongside when quoting GDT.
• Color convention: 上涨 (price up YoY) → red; 下跌 → green — Chinese stock-market convention. Just set the sign in the change string and the renderer handles the rest.
• No external chart libraries — tables with directional pills are enough; if a section has a time series, you may add a list of (date, value) tuples under a sparkline key and the renderer will draw an inline SVG-like bar.
• Do not add PageBreak markers between sections. The renderer takes care of pagination.
• Do not wrap sections in KeepInFrame mode=shrink — that was an old attempt and produces cramped sections.

5. Render the PDF

bash

python scripts/render_pdf.py --data work/report_data.json --out "乳制品市场行情报告_202310.pdf"

The renderer uses reportlab + the built-in STSong-Light CJK font, so no Chromium / Playwright / wkhtmltopdf installation is required. If reportlab is missing, install it with pip install reportlab.

Expected page count for a full 12-section report: 8–10 pages (cover + body). If you see 13+ pages, you have added a per-section page break somewhere — remove it.

6. Present results

Use the present_files tool to show the PDF (it previews as an artifact card). In the chat reply, give a 3–4 line summary of which sections came from which input PDFs and one highlighted takeaway from section 11.

Conventions & guardrails

• Numbers over vibes. Every claim in the report must be backed by a number from an input PDF, an official source in data-sources.md, or the GDT calculator. If a number is genuinely missing, write "数据待补充" — do not invent.
• Cite inline. Each section ends with a "数据来源" footer (e.g. "数据来源: 农业农村部; 艾格农业《中国乳业研究月报》202310").
• No fabrication of future events. Section 10 (走势预测) is forward-looking, but every prediction must be conditioned on a stated driver ("if GDT 投放量 continues to rise…"); do not give a single point estimate without a driver.
• Honor user's formulas exactly. GDT-to-warehouse formulas in section 8 are:
  • AMF: (contract2_price + 165) * FX * 1.13 + 300
  • WMP / SMP: (contract2_price + 115) * FX * 1.13 + 300
  • Contract-2 is the main contract, delivery 1–2 months out. If the PDF reports a different contract period, call it out.
• Stay in Chinese for prose. Allow English/Latin acronyms (GDT, AMF, WMP, SMP, USDA, NZX, AHDB, Fonterra).
• One report per period. If the user has multiple monthlies, generate one consolidated report for the latest month (with comparison vs. prior month) unless they ask for a multi-month roll-up.
• Layout discipline. No per-section PageBreak. No KeepInFrame shrink-mode per section. The renderer already produces a balanced 8–10 page magazine layout. If your output is 13+ pages, you have re-introduced one of those — fix the data builder, not the renderer.
• Color discipline. Use the bake-in tokens (NAVY / ACCENT / SOFT_BG / RED / GREEN) — do not invent new colors. Pill color is auto-driven by the +/- sign in the change string, so you do not pick the color manually.

Quick-start (minimal example)

text

User: "帮我跑一下 10 月的乳业月报"
→ Skill: detects 艾格农业-中国乳业研究月报202310.pdf
→ Extracts text → fills 12 sections → assembles data dict (per references/data-schema.md)
→ Calls scripts/render_pdf.py --data work/report_data.json --out 乳制品市场行情报告_202310.pdf
→ Presents: 乳制品市场行情报告_202310.pdf (8–10 pages, magazine style)

See references/report-template.md for the authoritative section spec, and references/data-schema.md for the data JSON schema the renderer expects.