PPTX Official

PPTX Skill

An Apache-2.0 toolkit for producing, editing, and reading Microsoft PowerPoint (.pptx) files. Written from scratch against the public ECMA-376 / ISO/IEC 29500 (PresentationML) specification and built on permissively-licensed tooling (python-pptx MIT, pptxgenjs MIT, lxml BSD-3-Clause, Pillow MIT-CMU, optional external binaries soffice MPL 2.0 and pdftoppm GPL) so it can be reused in commercial projects without restriction.

Decision matrix

If the task mixes several of these, do them in this order: read → plan slide-by-slide → edit/create → validate → visual QA.

One-time environment setup

DSH environment note: the original text targets MiMo's bundled runtime (MIMO_PYTHON / MIMO_SOFFICE / MIMO_NODE / MIMO_NODE_MODULES). DSH has no such runtime — use the system python or uv run for Python; install LibreOffice yourself for soffice; for the PptxGenJS authoring path run npm install pptxgenjs react react-dom sharp react-icons mathjax-full yourself. Replace every "$MIMO_PYTHON" with python.

Prerequisites

If uv or bun are not yet installed:

# Install uv (Python package/project manager)
# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
# Windows: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

# Install bun (TypeScript runtime, replaces Node.js for this workflow)
# macOS / Linux
curl -fsSL https://bun.sh/install | bash
# Windows: powershell -c "irm bun.sh/install.ps1|iex"

Only if the user explicitly refuses uv / bun, substitute pip (in a venv you manage yourself) for uv, and npm/pnpm + npx tsx for bun — everything else in this skill stays the same.

Python (uv)

Python dependencies are managed by uv. Do not use pip directly.

# Initialize project (if no pyproject.toml exists)
uv init -p 3.12

# Add dependencies
uv add python-pptx lxml Pillow
uv add defusedxml                  # safe XML parsing (recommended for manual XML edits)

Rules:

• Never use pip — always uv add for packages.
• Never run python scripts/... directly — always uv run scripts/....
• Don't manually manage environments with python -m venv or source .venv/bin/activate.

TypeScript (bun)

For PptxGenJS creation, use bun (project-local, not global installs):

# Initialize (if no package.json exists)
bun init -y

# Add dependencies
bun add pptxgenjs                  # core PPTX creation library
bun add react react-dom sharp      # rasterization (icons + formulas)
bun add react-icons                # icon library (FA, MD, etc.)
bun add mathjax-full               # LaTeX formula rendering

# Type definitions (including Bun runtime types)
bun add -d @types/bun @types/react @types/react-dom

Create a tsconfig.json if one doesn't exist:

{
  "compilerOptions": {
    "lib": ["ESNext"],
    "target": "ESNext",
    "module": "Preserve",
    "moduleResolution": "bundler",
    "allowImportingTsExtensions": true,
    "noEmit": true,
    "jsx": "react-jsx",
    "strict": true,
    "skipLibCheck": true,
    "types": ["bun"]
  }
}

Run scripts directly as TypeScript — no transpilation needed:

bun run create-ppt.ts

Always run type checking after writing or modifying TS code:

bun tsc --noEmit

Models may inadvertently use outdated PptxGenJS API signatures or deprecated syntax without realizing it. A type check catches these mismatches before runtime.

System dependencies (PDF/PNG rendering)

# macOS
brew install --cask libreoffice
brew install poppler

# Debian/Ubuntu
sudo apt-get install -y libreoffice poppler-utils

Every script under scripts/ uses only the standard library plus python-pptx, lxml, and Pillow. No proprietary dependencies. External binaries (soffice, pdftoppm) are invoked as subprocesses; nothing is bundled or statically linked.

Common commands

# 1. Extract text from every slide (title, body, notes) — the "what does it say?" query
uv run scripts/dump_text.py input.pptx --notes > input.txt

# 2. Convert a deck to PDF for review
uv run scripts/render_pdf.py input.pptx                     # writes input.pdf next to it

# 3. Convert every slide to a PNG (visual QA)
uv run scripts/render_slides.py input.pptx --out slides/       # writes slides/slide-1.png, ...

# 4. Grid thumbnail preview (planning which template slide to reuse)
uv run scripts/contact_sheet.py input.pptx --cols 3         # writes input.contact-sheet.jpg

# 5. Explode a .pptx into readable XML for surgical edits
uv run scripts/explode.py input.pptx unpacked/

# 6. Reassemble an exploded tree
uv run scripts/assemble.py unpacked/ output.pptx

# 7. Drop orphaned slides and unused media before reassembly
uv run scripts/prune.py unpacked/

# 8. Duplicate slide 3, or spin up a new slide from layout 5
uv run scripts/insert_slide.py unpacked/ --clone slide3.xml
uv run scripts/insert_slide.py unpacked/ --blank-from slideLayout5.xml

# 9. Well-formedness check (ZIP + XML + python-pptx round-trip)
uv run scripts/diagnose.py output.pptx

Every script is a small Python CLI. Some scripts (e.g. render_pdf.py, render_slides.py, contact_sheet.py) import from a shared helper (scripts/soffice_bridge.py) — copy the whole scripts/ directory together. Read the top of each file for its full CLI options.

Live preview

A live preview server is available for real-time slide feedback. Not started by default. When multi-slide work begins, ask the user if they want live preview enabled.

Only offer this in a pure command-line environment. This server is for the MiMoCode CLI. If you are running inside a host that embeds MiMoCode via the SDK — a web UI, a desktop app, an IDE plugin, etc. — that host almost certainly has its own native preview / file-open mechanism; use it instead and do NOT start this server.

If yes, and this is a CLI environment:

# `scripts/preview.ts` lives in this skill's bundle directory, not in
# your cwd. Prefix it with the absolute path shown in this skill's
# location header (the folder that contains SKILL.md) — refer to it
# as <SKILL_DIR> below.

# Start (spawns background server, prints URL, exits immediately)
bun run <SKILL_DIR>/scripts/preview.ts /path/to/output.pptx
bun run <SKILL_DIR>/scripts/preview.ts /path/to/output.pptx --port 5000

# Stop
bun run <SKILL_DIR>/scripts/preview.ts --stop /path/to/output.pptx

If the server is already running, preview.ts detects this via PID file and prints the existing URL instead of spawning a duplicate.

The background server watches the .pptx file, debounces (800ms), converts to PDF via LibreOffice, and pushes a WebSocket reload to the browser. The browser's native PDF viewer provides scroll, thumbnails, zoom, and search. Preview output goes to .pptx-preview/ (separate from qa/, no conflict with other scripts).

Give the user the printed URL to open in their browser.

Live preview is for humans only — it does NOT replace visual QA. The preview server lets the user watch progress. You must still run the visual QA subagent (render PNGs to qa/, spawn a vision model to inspect them) as described in the "Visual QA execution model" section. These are independent workflows:

• Preview server → user sees live updates in browser
• Visual QA subagent → automated inspection, catches issues you can't

When to regenerate the .pptx during multi-slide work: If the user has the preview server running, regenerate the .pptx (re-run the creation script) after completing each logical module — e.g. after finishing a section's slides, not after every single shape placement. This gives the user meaningful visual checkpoints without excessive intermediate renders.

Authoring principles

Slides are a visual surface. Users read the deck at 40 feet from the back of a room, or in a browser tab three inches wide on a phone. Both have to work. Keep these in mind:

1. One idea per slide. If you can't summarize the slide in a five-word title, split it in two. Long-form reasoning belongs in the doc that accompanies the deck, not on the slide itself.
2. Title, not label. The title is the thesis of the slide. "Revenue" is a label. "Revenue grew 34% on 22% headcount" is a title. Titles carry the argument; bodies carry the evidence.
3. Every slide earns its visuals. A slide without a chart, image, icon, or shape is usually a bullet dump. Turn it into a comparison, a stat callout, a diagram, or delete it. A "visual" is not necessarily a picture — a stat callout, a comparison shape, or a well-typeset quote is already a visual. When you do need a picture, see Image sourcing below.
4. Layouts, not per-slide geometry. For anything reused (section dividers, content pages, quote slides), define a slide_layout once and apply it. This makes swapping the theme a one-line change instead of a fifty-slide sweep.
5. Aspect ratio matches the target. 16:9 (default) for laptops and projectors; 4:3 only when explicitly asked (still common in academia and some corporate templates); 16:10 for older widescreen.
6. Speaker notes carry the words. Put the full narration into slide.notes_slide.notes_text_frame so the presenter can rehearse from the deck itself. On the slide, keep it to the phrase they can hold in their head.
7. Bullets are not the default. Bullet lists are the failure mode of most decks — comparisons, tables, icons-with-labels, and stat callouts almost always land better.

Image sourcing (choose the right channel)

"Every slide earns its visuals" does not mean "generate an image for every slide." Choose the source based on what the image does. Four channels, ordered by preference (lower cost + higher stability first):

L1 · Draw in code. Icons via react-icons / iconify. Charts via matplotlib / plotly / echarts. Flowcharts, comparisons, org charts via shapes + lines. Anything that is a data or concept visualization — never fetch or generate an image for this.

L2 · Search a stock library, then download the bytes. When a slide needs a generic real photo (city skyline, office desk, team collaboration, nature, stock imagery), use web_search with a site:unsplash.com / site:pexels.com / site:pixabay.com query to find a real URL, then download the image to a local file and pass that path to add_picture (see Downloading below). Do NOT try to pass an HTTP URL directly to add_picture — python-pptx's add_picture only accepts a local path or a file-like object; it will NOT fetch a URL for you.

L3 · Search a specific source. For specific real things (a particular company's logo, a specific product's official screenshot, a named person's photo), use web_search with a targeted query (e.g. "Acme Inc" logo site:acme.com, or <product name> screenshot) instead of a stock-library query, then download the bytes the same way as L2. Never call image_gen here — a generated logo will not look like the real logo. Note: tavily_extract returns text/markdown/html only and cannot deliver binary image bytes — use bash + curl or python urllib to download.

L4 · Generate with image_gen. Only for stylized visuals — cover art, hero backgrounds, illustration-style concept images, brand-mood imagery. Budget: at most 1–2 image_gen calls per deck, typically for cover or section dividers. Not for every content slide.

Downloading a URL for add_picture

Two working patterns. Both keep the file local so add_picture can read the bytes:

# Pattern A: download to a temp file, then pass the path
from urllib.request import Request, urlopen
from pathlib import Path

def download_image(url: str, dest: Path) -> Path:
    req = Request(url, headers={"User-Agent": "Mozilla/5.0"})  # some CDNs 403 an empty UA
    with urlopen(req, timeout=15) as r:
        dest.write_bytes(r.read())
    return dest

path = download_image(hit_url, Path("assets/hero.jpg"))
slide.shapes.add_picture(str(path), Inches(1), Inches(1.5), width=Inches(11))

# Pattern B: in-memory via BytesIO — no temp file, but same request headers
from io import BytesIO
from urllib.request import Request, urlopen

req = Request(hit_url, headers={"User-Agent": "Mozilla/5.0"})
with urlopen(req, timeout=15) as r:
    slide.shapes.add_picture(BytesIO(r.read()), Inches(1), Inches(1.5), width=Inches(11))

Or from a bash step:

curl -sSL -A "Mozilla/5.0" -o assets/hero.jpg "$URL"

Wrap any download in a try/except: on failure fall through to the next channel (L3 → L4) or a shape+text fallback — never leave a slide blank.

Decision (run per slide)

1. Does this slide actually need a picture? Often the answer is no — a large stat callout, a comparison shape, a well-typeset quote, or a diagram is a visual. Skip pictures when layout + typography carries the message.
2. If yes, pick the cheapest channel that works:
   • Data / concept / flow / icon → L1 (draw in code)
   • Generic real photo → L2 (stock library search + download)
   • Specific brand / logo / product / person → L3 (targeted web search + download)
   • Stylized cover / hero / illustration → L4 (image_gen)
3. If a fetch fails: L2 → L3 → L4 → shape + text fallback. Never leave a slide blank because an image failed to load.

Anti-patterns

• Calling image_gen on every content slide. Slow, expensive, style-inconsistent, visually noisy. A 20-slide deck with 20 generated images is a red flag, not a success.
• Passing an HTTP URL to add_picture. python-pptx will raise FileNotFoundError — always download the bytes first (see Downloading).
• Using tavily_extract to grab image bytes. tavily_extract returns text only. Use bash + curl or python urllib for binaries.
• Making up a stock-library URL from memory. Unsplash / Pexels CDN paths are opaque hashes — you cannot reliably recall a URL that both exists AND matches the described content. Always search first, then download.
• Generating a logo, celebrity, or product screenshot with image_gen. The output will not resemble the real thing. Use targeted web search (L3).
• Executive / status / weekly decks in illustration style. Work reporting is data + icons + hierarchy, not concept art. Reserve L4 for launch, brand, or hero visuals.
• Leaving a slide blank because an image fetch failed. Always have a shape + text fallback; a well-formatted stat callout is a better slide than an empty one anyway.

Scene defaults (rough mix per deck type)

L2/L3 both cost about one web_search + one download per image — cheap compared to L4, still not free. L1 remains the default.

Typography defaults (safe starting point)

Change the palette for the topic (financial → navy #1F3A5F; environment → forest #2C5F2D; product launches → your brand's accent). Avoid pure black on pure white for backgrounds — #F7F5F0 cream on #1F1F1F ink reads softer under a projector.

Slide sizes

QA checklist — always run before declaring done

Assume something is wrong. PowerPoint opens broken files quietly: a misaligned text box, a chart pointing at deleted data, a stray placeholder that survived template fill. Verify explicitly.

1. Opens cleanly. No repair dialog, no missing-part warning.

   uv run scripts/diagnose.py output.pptx
   

2. Text integrity. No placeholder residue and no unfilled {{token}}s:

   uv run scripts/dump_text.py output.pptx --notes \
       | grep -Ei "\{\{|TODO|TBD|lorem|ipsum|xxxx|click to add"
   

   Grep must return nothing.

3. Visual sanity. Render the whole deck to PNG, spot-check the first, last, and any slide you touched. Look for:

   • Text overflowing the placeholder or getting auto-shrunk past readability.
   • Two shapes overlapping (title over image, footer over content).
   • Legend / axis labels cut off on charts.
   • Icons at the wrong scale (tiny hairline icons, or huge stretched ones).
   • Off-brand colors that snuck in from a copied slide.

   uv run scripts/render_slides.py output.pptx --out qa/
   

4. Layout hygiene. Every non-master slide should reference a real layout, not slideLayout1 by default for a section divider:

   uv run python -c "
   from pptx import Presentation
   prs = Presentation('output.pptx')
   for i, s in enumerate(prs.slides, 1):
       print(f'slide {i}: layout={s.slide_layout.name!r}')"
   

If any of these fail, fix and re-run — don't paper over.

Visual QA execution model

Slide images are expensive. A single rendered PNG at 150 DPI consumes thousands of context tokens. Loading multiple slides into the main conversation for inspection will quickly exhaust your context budget and crowd out useful working memory.

Default: always use a subagent for visual inspection. Spawn a general subagent with the rendered PNG paths and the inspection criteria from step 3 above. The subagent reports findings as text (slide number + issue description); the images never enter the main conversation context. This is mandatory unless the exception below applies.

actor({
  operation: {
    action: "run",
    subagent_type: "general",
    // omit `model` when your current model is vision-capable (preferred — see Model selection)
    description: "Visual QA slides",
    prompt: "Inspect the rendered slide images in qa/ for: text overflow, overlapping shapes, cut-off labels, wrong-scale icons, off-brand colors. Report each issue as 'slide N: <problem>'. Images: qa/slide-1.png through qa/slide-<N>.png."
  }
})

Model selection (in priority order):

1. Prefer the user's current model. Query the servable vision models via actor({ operation: { action: "models", vision: true } }). If your current model is in the list, omit the model parameter — the subagent inherits it, and visual QA runs on the model the user chose.
2. Fallback: lightweight vision model, stated explicitly. If your current model is not vision-capable, pick a lightweight vision model from the query result (haiku/flash/lite-class if present, otherwise the cheapest listed) and tell the user in your reply which model you used for visual QA and why (current model cannot inspect images).
3. No vision model available. If the query returns an empty list, do not guess a model id — skip the visual QA subagent, run the structural checks only, and tell the user visual inspection could not be performed.

Never hardcode a model id. Which models are servable varies per deployment and changes over time; only ids returned by the vision models query are guaranteed to work. A guessed id fails the subagent outright.

Exception — direct inspection in the main context: Only load slide images directly (without a subagent) when the user explicitly requests that the current model inspect a specific slide for fine-grained, interactive editing (e.g. "look at slide 5 and adjust the title position"). This requires the current model to be multimodal. If it isn't, inform the user and offer to spawn a vision subagent instead.

Common visual pitfalls

• Titles wrap onto two lines — either shorten the title or widen the placeholder. Wrapped titles push body content down and break layout alignment across slides.
• Body text auto-shrinks below 14pt — python-pptx respects the placeholder's autofit setting; if the resulting size is unreadable, split the slide instead of accepting the shrink.
• Charts inherit Office defaults — the pale blue / gray palette shipped with PowerPoint looks generic. Explicitly set chart.chart_style or use python-pptx's low-level access to set fill colors on series.
• Speaker notes forgotten — a deck without notes cannot be rehearsed. Fill notes_slide.notes_text_frame.text on every slide, even if just a single sentence.
• Images at wrong DPI — a 4000×3000 photo on a 1280×720 slide bloats the file with no visual benefit. Resample down to ~150 DPI at the target display size (see create.md → Images).
• Fonts not embedded — python-pptx does not embed fonts. If the deck is opened on a machine without the chosen font, PowerPoint substitutes, and layout drifts. For Latin text, prefer system-safe fonts (Calibri, Arial, Segoe UI, Times New Roman, Consolas) or ship the .pptx alongside a font install step.
• CJK text needs the East-Asian font slot — run.font.name only sets the Latin typeface (a:latin); Chinese / Japanese / Korean glyphs come from the East-Asian slot (a:ea), which python-pptx does not expose. Leave it unset and CJK renders as tofu boxes or an inconsistent substitute. Set a:latin + a:ea + a:cs to a CJK-capable font on every run that contains CJK text (recipe in create.md → CJK / East-Asian text).

What is out of scope

• .ppt (PowerPoint 97-2003 binary). Convert first: soffice --headless --convert-to pptx old.ppt.
• VBA / macros / .pptm. This skill does not emit or execute macros.
• Password-protected or encrypted decks. python-pptx cannot read encrypted files; strip protection with PowerPoint or LibreOffice first.
• Live PowerPoint automation. For COM (Windows) or AppleScript (macOS) integration, use a dedicated automation library — this toolkit is file-in / file-out.
• Keynote .key files. Not a PresentationML format; use Apple's Keynote or LibreOffice for round-trip.

Where each detail lives

• Creating from scratch: create.md — python-pptx recipes, PptxGenJS recipes, layouts, text, tables, images, charts, icons, backgrounds, speaker notes, palette and typography guidance.
• Editing / templating: edit.md — placeholder fill, slide duplication / reorder / delete, explode/assemble for XML surgery, comments, cleanup of orphaned parts, common pitfalls.
• Reading / extracting: read.md — plain-text export (including speaker notes), structural walk, metadata, thumbnails, image extraction, conversion to PDF / PNG for QA.
• Scripts: scripts/ — CLI utilities (some share a local soffice_bridge.py helper; copy together when extracting).
• Live preview: scripts/preview.ts — launcher (start/stop); scripts/preview_server.ts — background server that watches .pptx, converts to PDF, serves with WebSocket hot-reload in the browser's native PDF viewer.