documents-docx-skill

name	documents-docx-skill
description	Use when generating Word documents (.docx). Uses the marble_docx API.

Python-docx Skill: DOCX Document Generation

Marble Library (PREFERRED)

Use the marble_docx helper library. It provides themed headings, tables, rules, cover pages, two-column layouts, and paragraphs — all with automatic theming.

IMPORTANT: Always use the library helpers below. Never manually define color constants, helper functions, or cell shading logic — the library handles all of this.

This skill is the complete API reference — all palettes, font pairs, and methods are documented below. Use them directly in your code without any exploration calls.

from marble_docx import Doc

doc = Doc(palette='ocean', fonts='modern')
# palette: midnight, forest, coral, terracotta, ocean, charcoal, teal, berry, sage, cherry
# fonts: classic, modern, bold, elegant, clean
# Or: Doc() for default light theme, Doc('dark') for dark theme

API Quick Reference

Method	Purpose
Doc(theme, font, palette, fonts)	Create document. palette for named colors, fonts for font pair
doc.cover_page(title, subtitle, author, date)	Professional title page with accent lines
doc.toc_placeholder()	Table of Contents heading placeholder
doc.heading(text, level)	Heading with primary accent. Level 0=Title, 1=H1, 2=H2
doc.rule(color)	Horizontal divider line (default: primary). Use after section headings
doc.paragraph(text, bold, italic, color, size, align)	Styled paragraph. Supports <b> and <i>
doc.caption(text)	Secondary italic note
doc.rich(*parts)	Mixed-style paragraph. Parts: strings or {'text', 'bold', 'color', ...} dicts
doc.bullet(text)	Bullet list item. Supports <b> and <i>
doc.numbered(text)	Numbered list item
doc.table(headers, rows, col_widths, header_bg, header_fg)	Themed table with alternating rows
doc.two_column_table(left_rows, right_rows)	Side-by-side content via invisible table
doc.image(path, width, align)	Insert image
doc.page_break()	Page break
doc.header(text, align, color, bold)	Set page header
doc.footer(text, align, color, bold)	Set page footer
doc.raw	Access underlying docx.Document for advanced ops

Named Palettes

Choose a palette that matches the content's mood — not the default blue:

Palette	Primary	Best for
midnight	Deep navy	Finance, enterprise, formal
forest	Rich green	Sustainability, health, growth
coral	Warm pink-red	Marketing, creative, consumer
terracotta	Earthy brown	Luxury, real estate, hospitality
ocean	Deep blue	Tech, analytics, professional
charcoal	Dark gray	Minimal, corporate, neutral
teal	Vibrant teal	Modern, SaaS, healthcare
berry	Deep plum	Premium, beauty, fashion
sage	Muted green-blue	Consulting, calm, wellness
cherry	Bold red	Bold reports, urgency, food

Font Pairs

Name	Heading / Body
classic	Georgia / Calibri
modern	Calibri / Calibri Light
bold	Arial Black / Arial
elegant	Cambria / Calibri
clean	Trebuchet MS / Calibri

Design Guidelines

1. Document structure

For reports and long documents, follow this structure:

1. Cover page — doc.cover_page() + doc.page_break()
2. Table of Contents — doc.toc_placeholder() + doc.page_break() (optional)
3. Executive Summary — heading + key findings
4. Body sections — each with heading + rule + content
5. Appendix — data tables, methodology notes

2. Use doc.rule() after every section heading

This gives visual structure and separates sections clearly:

doc.heading('1. Executive Summary', level=1)
doc.rule()
doc.paragraph('Key findings...')

3. Always set col_widths on tables

Without explicit widths, columns render unevenly:

doc.table(headers, rows, col_widths=[3.5, 1.5, 1.5])  # GOOD
doc.table(headers, rows)  # BAD — uneven columns

4. Use rich() for mixed-style inline content

doc.rich(
    'Total spend was ',
    {'text': '$793', 'bold': True},
    ' with a ROAS of ',
    {'text': '0.70x', 'bold': True, 'color': 'error'},
)

5. Use captions for attributions and notes

doc.caption('Source: Meta Ads API | Data as of Feb 23, 2026')

6. Color-coded table cells

{'text': '1.55x', 'color': 'success', 'bold': True}  # green for positive
{'text': '0.38x', 'color': 'error', 'bold': True}     # red for negative

Bold header variant: header_bg='primary', header_fg='text_on_primary'

7. Embedding ad creatives, thumbnails, and other platform-fetched images

When the document needs real images returned by previous tool calls (ad thumbnails from facebook_analyze_ad_creative_by_id_or_url, competitor ad images from the ad library, etc.), pass each image URL as an attachment_url data source on the SAME advanced_analysis call. The runtime downloads them into the workspace before your code runs — doc.image() takes the local path:

# Tool call payload:
# dataSources: [
#   {type: 'attachment_url', identifier: 'https://cdn.fbcdn.net/.../thumb_a.jpg', alias: 'creative_1'},
#   {type: 'attachment_url', identifier: 'https://cdn.fbcdn.net/.../thumb_b.png', alias: 'creative_2'},
# ]

# In code:
from marble_docx import Doc
doc = Doc(palette='ocean')
doc.heading('Top Performers', level=1)
doc.rule()
doc.image(f'{WORKSPACE}/creative_1.jpg', width=4)
doc.caption('Figure 1: Outdoor Adventure (Video) — CTR 2.46%')
doc.image(f'{WORKSPACE}/creative_2.png', width=4)
doc.caption('Figure 2: Forest Boots — CTR 1.92%')
doc.save(f'{WORKSPACE}/report.docx')

Do NOT import requests/urllib from inside your code — network is blocked and the call will fail with DNS errors. The attachment_url mechanism is the supported path. Extensions (.jpg/.png/.webp/.gif etc.) are auto-detected.

Anti-Patterns — DO NOT

1. NEVER use \n in text — always call doc.paragraph() or doc.bullet() separately for each line.
2. Don't dump raw data — summarize with tables + bullet highlights. Max 6-8 columns per table.
3. Don't use giant tables — if you have 10+ columns, transpose the table or split into multiple tables.
4. Don't forget col_widths — every table needs explicit column widths.
5. Don't skip cover pages for formal reports — use doc.cover_page() for anything over 2 pages.
6. Don't use doc.heading() without doc.rule() — headings need visual separation.
7. Don't manually define colors — use theme keys: 'primary', 'success', 'error', 'secondary'.
8. Don't use page_break() after every section — only use between major sections (e.g. after cover page, before appendix). Excessive page breaks create blank pages.
9. NEVER use raw python-docx enums without importing them — if using doc.raw, always import what you need: from docx.enum.table import WD_TABLE_ALIGNMENT, from docx.enum.text import WD_ALIGN_PARAGRAPH, from docx.shared import Pt, Inches, RGBColor. The marble_docx library handles this automatically — prefer it over raw API.
10. Respect page count constraints — if the user asks for "a 2-page report", the final document must be ~2 content pages (cover page is extra). Plan your layout BEFORE writing code: pick 2-3 key sections max, use compact tables, limit charts to 1-2, and skip exhaustive breakdowns. A page holds ~4-5 sections of heading+paragraph, OR 1 table + 2 paragraphs + 1 chart. Fewer page_break() calls = fewer pages.

Example — Professional Report

from marble_docx import Doc

doc = Doc(palette='ocean', fonts='elegant')

doc.header('B&M Advertising  |  Performance Report', align='center', color='primary', bold=True)
doc.footer('Powered by GoMarble AI  |  Feb 2026')

# ── Cover page ──
doc.cover_page(
    'B&M Advertising Account',
    subtitle='Spend & ROAS Performance Report',
    author='GoMarble AI',
    date='February 17 – February 23, 2026',
)
doc.page_break()

# ── Executive Summary ──
doc.heading('1. Executive Summary', level=1)
doc.rule()
doc.paragraph('This report covers Meta Ads performance for B&M Advertising over the past 7 days.')
doc.rich(
    'Total spend was ',
    {'text': '$793.95', 'bold': True},
    ' with a blended ROAS of ',
    {'text': '0.70x', 'bold': True, 'color': 'error'},
    ' — below the 1.0x break-even threshold.',
)
doc.caption('Note: Weighted ROAS is the spend-weighted average across all campaigns.')

# ── Key Metrics (two-column) ──
doc.heading('2. Key Metrics', level=1)
doc.rule()
doc.two_column_table(
    ['Total Spend: $793.95', 'Active Campaigns: 4'],
    ['Blended ROAS: 0.70x', 'Top ROAS: 2.83x (Test)'],
)

# ── Campaign Breakdown ──
doc.heading('3. Campaign Performance', level=1)
doc.rule()
doc.table(
    headers=['Campaign', 'Spend', 'ROAS', 'Status'],
    rows=[
        ['Prospecting', '$320', {'text': '0.55x', 'color': 'error', 'bold': True}, 'Active'],
        ['ASC', '$200', {'text': '0.38x', 'color': 'error', 'bold': True}, 'Active'],
        ['BOF', '$137', {'text': '1.55x', 'color': 'success', 'bold': True}, 'Active'],
        ['Test', '$136', {'text': '2.83x', 'color': 'success', 'bold': True}, 'Active'],
    ],
    col_widths=[2.5, 1.2, 1.2, 1.2],
    header_bg='primary', header_fg='text_on_primary',
)

# ── Chart ──
doc.image(f'{WORKSPACE}/roas_chart.png', width=6)
doc.caption('Figure 1: ROAS by Campaign')

# ── Recommendations ──
doc.heading('4. Recommendations', level=1)
doc.rule()
doc.bullet('<b>Pause Prospecting</b> — 0.55x ROAS, reallocate budget to BOF.')
doc.bullet('<b>Scale BOF</b> — 1.55x ROAS with $137 spend, room to grow.')
doc.bullet('<b>Expand Testing</b> — 2.83x ROAS is the highest performer.')
doc.bullet('<b>Review ASC targeting</b> — 0.38x ROAS suggests audience/creative issues.')

doc.save(f'{WORKSPACE}/report.docx')
result = {"title": "B&M Performance Report", "description": "Multi-page Word document with cover page and recommendations."}

Raw API (fallback)

For anything the library doesn't cover, use doc.raw:

doc = Doc()
raw = doc.raw  # docx.Document object

# Helpers also exported:
from marble_docx import shade_cell, replace_image
shade_cell(cell, '1A73E8')
replace_image(doc, 0, f'{WORKSPACE}/new_chart.png', width=5.0)

Editing an Existing DOCX

For edits (fix text, update a table, remove a section), open the existing file — do NOT rebuild from scratch.

from docx import Document

doc = Document(f'{WORKSPACE}/report.docx')

Navigate paragraphs

for i, para in enumerate(doc.paragraphs):
    if para.text.strip():
        print(i, para.style.name, para.text[:80])

Change text in a paragraph

para = doc.paragraphs[5]
for run in para.runs:
    run.text = run.text.replace('old value', 'new value')

Replace an entire paragraph's text

para = doc.paragraphs[5]
para.clear()
run = para.add_run('New paragraph text')
run.font.size = Pt(11)

Delete a paragraph

para = doc.paragraphs[5]
p = para._element
p.getparent().remove(p)

Modify a table cell

table = doc.tables[0]  # first table
table.cell(1, 2).text = '$1,250'  # row 1, col 2

Add a row to an existing table

table = doc.tables[0]
row = table.add_row()
row.cells[0].text = 'New Campaign'
row.cells[1].text = '$500'

Replace an image (e.g. regenerate a chart at different size)

from marble_docx import replace_image

doc = Document(f'{WORKSPACE}/report.docx')
# List images to find the right index
for i, shape in enumerate(doc.inline_shapes):
    print(i, shape.width, shape.height)
# Replace image 0 with a new file, optionally resize
replace_image(doc, 0, f'{WORKSPACE}/new_chart.png', width=5.0)
doc.save(f'{WORKSPACE}/report.docx')

Save back

doc.save(f'{WORKSPACE}/report.docx')
result = {"title": "Updated Report", "description": "Fixed table values"}

When to edit vs rebuild

• Edit: fix text, update table cells, delete a paragraph, change formatting, replace an image — use Document(path)
• Rebuild section: if a whole section is wrong, delete those paragraphs and add new ones
• Full rebuild: only if the entire structure needs to change

Saving

Always save to WORKSPACE:

doc.save(f'{WORKSPACE}/filename.docx')