| name | professional-document |
| description | Produce a formal business document as .docx (python-docx) or PDF (reportlab) - reports, proposals, briefings, one-pagers, meeting summaries. Use whenever the task asks for a written document, report file, proposal, or PDF. Covers structure, typography, build and delivery. |
Professional Document Builder
Produce a business document that reads and looks like it was prepared by a
professional analyst: clear structure, restrained typography, real numbers.
Language rule
All user-visible text in the document MUST be written in the same language
as the task description you received. This skill's instructions are English,
but the deliverable follows the task's language.
Format choice
.docx (python-docx) when the reader may edit: proposals, drafts,
working reports. Default choice when the task does not specify.- PDF (reportlab,
platypus flowables) when the task says PDF or the
document is final/customer-facing: signed proposals, formal briefings,
one-pagers.
- Do NOT use weasyprint or LibreOffice conversion (system dependencies are
not guaranteed). reportlab and python-docx are pure-python and reliable.
- CJK note (PDF only): reportlab's built-in fonts cannot render Japanese,
Chinese or Korean. For CJK text register a CID font first:
from reportlab.pdfbase import pdfmetrics; from reportlab.pdfbase.cidfonts import UnicodeCIDFont; pdfmetrics.registerFont(UnicodeCIDFont('HeiseiKakuGo-W5'))
and set that font on every style. python-docx handles CJK natively.
Process
- python-docx, reportlab and matplotlib are usually preinstalled by the
environment warm-up - try the import first; pip install only if missing.
- Draft the full document content in markdown first (
draft.md). Get the
substance right before any formatting code.
- Convert the draft to the target format with a build script.
- Verify (below), fix, rebuild.
- Deliver (below).
Structure
Standard business report skeleton - adapt, do not skip levels:
- Title block: document title, subtitle/purpose, date, author line
(the requesting organization, not you). Documents longer than ~6 pages
get a cover page plus a table of contents built from the headings.
- Executive summary: 3-6 sentences a busy executive could read alone -
the key finding, the number that matters, the recommendation.
- Body sections (3-6): each opens with a one-sentence takeaway in bold,
followed by supporting analysis, tables and figures.
- Recommendations / next steps: numbered, concrete, each with an owner
or function and a timeframe where possible.
- Appendix (optional): methodology, data notes, detailed tables.
One-pagers compress the same skeleton to: title block, summary paragraph,
3-4 key points with numbers, recommendation box.
Typography and layout
- One font family throughout. Body 10.5-11pt, line spacing 1.15-1.3.
- Heading hierarchy: H1 16-18pt bold, H2 13-14pt bold, H3 11pt bold.
Use color
#1A2B4A for headings, near-black #222222 for body.
- Margins: 2.0-2.5 cm all around. Page numbers in the footer from page 2.
- Numbers that carry the argument go in bold; do not bold whole sentences.
- Tables: bold header row with a light fill (
#F2F5FA), thin borders,
right-align numeric columns, thousands separators.
- Charts: matplotlib PNGs at 150+ dpi, sized to the text column width,
numbered captions below (Figure 1: ...). Same palette as the tables.
- Charts with non-Latin labels (Japanese etc.): matplotlib's default font
has no CJK glyphs (hollow boxes). Register a font first - for Japanese:
pip install japanize-matplotlib, then import japanize_matplotlib once.
For other scripts register a Noto Sans font via
matplotlib.font_manager.fontManager.addfont(). Verify the rendered PNG
has no hollow-box glyphs before embedding it.
- Never fake letterhead or signatures.
Content quality bar
- Every claim backed by the provided data or explicitly sourced web
research (cite source name + URL inline or in the appendix).
- Round numbers for readability; keep raw precision in appendix tables.
- No filler phrases. Cut any sentence that does not inform a decision.
Verification (mandatory)
- docx: re-open with python-docx; assert heading structure exists and all
planned sections are present and non-empty.
- PDF: re-open with
pypdf (usually preinstalled by warm-up); assert
expected page count (>= 1) and that extracted text contains the section
headings AND the CJK strings you wrote (garbled fonts extract as empty
or replacement characters - if a heading is missing, fix the font).
Fix and rebuild until clean. At least one full review-and-rebuild pass is
mandatory: re-read the built document against the task's stated quality
conditions before delivering.
Delivery
- Save with a descriptive filename, e.g.
supplier_risk_briefing.pdf.
- If the task provided an upload URL for deliverables, upload with:
curl -sS -X PUT --upload-file <file> -H "Content-Type: <mime>" "<upload_url>"
(docx mime: application/vnd.openxmlformats-officedocument.wordprocessingml.document,
pdf mime: application/pdf). Retry once on failure.
- In your final report, state the filename, page/section count, and
reproduce the executive summary verbatim.
