PowerPoint Skill

Generates, updates, and manages PowerPoint slide decks using python-pptx with YAML-driven content and styling definitions.

Overview

This skill provides Python scripts that consume YAML configuration files to produce PowerPoint slide decks. Each slide is defined by a content.yaml file describing its layout, text, and shapes. A style.yaml file defines dimensions, template configuration, layout mappings, metadata, and defaults.

SKILL.md covers technical reference: prerequisites, commands, script architecture, API constraints, and troubleshooting. For conventions and design rules (element positioning, visual quality, color and contrast, contextual styling), follow pptx.instructions.md.

Prerequisites

PowerShell

The Invoke-PptxPipeline.ps1 script handles virtual environment creation and dependency installation automatically via uv sync. Requires uv, Python 3.11+, and PowerShell 7+.

Installing uv

If uv is not installed, use a package manager that verifies the download against a published hash:

# Windows
winget install --id astral-sh.uv -e

# macOS / Linux (Homebrew)
brew install uv

To install from a release archive instead, download it from the uv releases page and check it against its .sha256 file before extracting.

System Dependencies (Export and Validation)

The Export and Validate actions require LibreOffice for PPTX-to-PDF conversion and optionally pdftoppm from poppler for PDF-to-JPG rendering. When pdftoppm is not available, PyMuPDF handles the image rendering.

The Validate action's vision-based checks require the GitHub Copilot CLI for model access.

# macOS
brew install --cask libreoffice
brew install poppler        # optional, provides pdftoppm

# Linux
sudo apt-get install libreoffice poppler-utils

# Windows (winget preferred, choco fallback)
winget install TheDocumentFoundation.LibreOffice
# choco install libreoffice-still      # alternative
# poppler: no winget package; use choco install poppler (optional, provides pdftoppm)

Copilot CLI (Vision Validation)

The validate_slides.py script uses the GitHub Copilot SDK to send slide images to vision-capable models. The Copilot CLI must be installed and authenticated:

# Install Copilot CLI
npm install -g @github/copilot-cli

# Authenticate (uses the same GitHub account as VS Code Copilot)
copilot auth login

# Verify
copilot --version

Required Files

• style.yaml — Dimensions, defaults, template configuration, and metadata
• content.yaml — Per-slide content definition (text, shapes, images, layout)
• (Optional) content-extra.py — Custom Python for complex slide drawings

Content Directory Structure

All slide content lives under the working directory's content/ folder:

content/
├── global/
│   ├── style.yaml              # Dimensions, defaults, template config, and theme metadata
│   └── voice-guide.md          # Voice and tone guidelines
├── slide-001/
│   ├── content.yaml            # Slide 1 content and layout
│   └── images/                 # Slide-specific images
│       ├── background.png
│       └── background.yaml     # Image metadata sidecar
├── slide-002/
│   ├── content.yaml            # Slide 2 content and layout
│   ├── content-extra.py        # Custom Python for complex drawings
│   └── images/
│       └── screenshot.png
├── slide-003/
│   ├── content.yaml
│   └── images/
│       ├── diagram.png
│       └── diagram.yaml
└── ...

Global Style Definition (style.yaml)

The global style.yaml defines dimensions, template configuration, layout mappings, metadata, and defaults. Color and font choices are specified per-element in each slide's content.yaml rather than centralized in the style file.

See the style.yaml template for the full template, field reference, and usage instructions.

Per-Slide Content Definition (content.yaml)

Each slide's content.yaml defines layout, text, shapes, and positioning. All position and size values are in inches. Color values use #RRGGBB hex format or @theme_name references.

Text contract: markdown-like list lines in textbox.text and shape.text are interpreted as PowerPoint lists during rendering. Unordered markers (-, +, *) become bulleted paragraphs, ordered markers (1., 1)) become auto-numbered paragraphs, and leading indentation maps to paragraph level.

See the content.yaml template for the full template, supported element types, supported shape types, and usage instructions.

Accessibility: each slide's title becomes the slide title that screen readers announce and navigate by, alt or decorative on an image sets its alternative text, and metadata.language in style.yaml sets the document and text language.

Complex Drawings (content-extra.py)

When a slide requires complex drawings that cannot be expressed through content.yaml element definitions, create a content-extra.py file in the slide folder. The render() function signature is fixed. The build script calls it after placing standard content.yaml elements.

See the content-extra.py template for the full template, function parameters, and usage guidelines.

Security Validation

content-extra.py execution is disabled by default. When a slide folder contains one and --allow-scripts is not passed, the build fails with an error naming the file. Pass --allow-scripts to authorize execution after reviewing the script.

When execution is authorized, the build script performs AST-based static analysis before running the file and rejects the patterns below. This analysis is a lint that catches obvious mistakes early. It is not a security boundary: a blocked module reached through an alias is not detected, and pathlib and open are permitted. Authorization is the control.

Allowed imports:

• pptx and all pptx.* submodules
• Safe standard-library modules (e.g., math, copy, json, re, pathlib, collections, itertools, functools, typing, enum, dataclasses, decimal, fractions, string, textwrap)

Blocked imports:

• subprocess, os, shutil, socket, ctypes, signal, multiprocessing, threading, http, urllib, ftplib, smtplib, imaplib, poplib, xmlrpc, webbrowser, code, codeop, compileall, py_compile, zipimport, pkgutil, runpy, ensurepip, venv, sqlite3, tempfile, shelve, dbm, pickle, marshal, importlib, sys, telnetlib
• Any third-party package not on the allowlist

Blocked builtins:

• Dangerous: eval, exec, __import__, compile, breakpoint
• Indirect bypass: getattr, setattr, delattr, globals, locals, vars
• Attribute-form calls onto builtins, os, sys, subprocess, and importlib (for example builtins.eval(...))

--allow-scripts flag:

Pass --allow-scripts to authorize execution of content-extra.py files. Without it, a present script fails the build rather than being skipped, so a deck never silently loses custom drawings. The flag authorizes execution; it does not skip the lint.

This is a behavior change. A deck that previously built with a content-extra.py present now requires --allow-scripts, and a script relying on a blocked import no longer has a bypass.

python scripts/build_deck.py \
  --content-dir content/ \
  --style content/global/style.yaml \
  --output slide-deck/presentation.pptx \
  --allow-scripts

When validation fails, the build raises ContentExtraError with a message identifying the violation and file path.

Script Reference

The full command surface lives in references/script-reference.md: build a deck, build from a template, update specific slides, extract content from an existing PPTX, validate, export slides to images or SVG, dry-run validation, generate theme variants, and embed audio.

Script Architecture

The build and extraction scripts use shared modules in the scripts/ directory:

python-pptx Constraints

• python-pptx does NOT support SVG images. Always convert to PNG via cairosvg or Pillow.
• python-pptx cannot create new slide masters or layouts programmatically. Use blank layouts or start from a template PPTX with the --template argument.
• Transitions and animations are preserved when opening and saving existing files, but cannot be created or modified via the API.
• When extracting content, slide master and layout inheritance means many text elements have no inline styling. Add explicit font properties in content YAML before rebuilding.
• The Export and Validate actions require LibreOffice for PPTX-to-PDF conversion. The PowerShell orchestrator checks for LibreOffice availability before starting and provides platform-specific install instructions if missing.
• Accessing background.fill on slides with inherited backgrounds replaces them with NoFill. Check slide.follow_master_background before accessing the fill property.
• Gradient fills use the python-pptx GradientFill API with GradientStop objects. Each stop specifies a position (0–100) and a color.
• Theme colors resolve via MSO_THEME_COLOR enum. Brightness adjustments apply through the color format's brightness property.
• Template-based builds load layouts by name or index. Layout name resolution falls back to index 6 (blank) when no match is found.

Security Considerations

This skill processes PDF files via PyMuPDF, which wraps the MuPDF C library. MuPDF parses untrusted binary structures (cross-reference tables, stream objects, font definitions) and historical CVEs have shown that memory-safety bugs in C parsers can lead to crashes, memory disclosure, or in rare cases code execution.

Mitigations in place

• scripts/pdf_safety.py validates every PDF (existence, regular-file, size <= 100 MB, %PDF- magic bytes, page count <= 1000) before calling fitz.open().
• All fitz operations are wrapped in safe_open_pdf(), which converts MuPDF exceptions into typed PdfSafetyError subclasses (PdfTooLargeError, PdfInvalidFormatError, PdfTooManyPagesError, PdfParseError, PdfRenderError).
• pymupdf is version-pinned in pyproject.toml (>=1.27.1,<2.0) to track security fixes without silently adopting a major-version API change.

Defense-in-depth note: the 5-byte %PDF- magic check is a necessary but not sufficient guarantee of structural validity. A crafted small file that begins with %PDF- still reaches the MuPDF parser, where memory-safety bugs may exist. Callers MUST keep PyMuPDF patched against the latest advisories (tracked via microsoft/hve-core#1020 / pip-audit) and continue to treat such inputs as untrusted. The PdfSafetyError hierarchy (PdfTooLargeError, PdfInvalidFormatError, PdfTooManyPagesError, PdfParseError, PdfRenderError) is one defense layer alongside the version pin and CVE monitoring; no single layer is sufficient on its own.

Accepted risk

These mitigations reduce but do not eliminate the C-extension attack surface. Consumers SHOULD treat PDF inputs from outside this skill's PPTX-to-PDF pipeline as untrusted and apply additional sandboxing (subprocess, container) before feeding adversarial input.

Update policy

Re-check NVD and OSV advisories for MuPDF and PyMuPDF quarterly and on every pip-audit alert.

Cross-references

• microsoft/hve-core#1018 — original hardening request
• microsoft/hve-core#1020 — pip-audit CI for ongoing CVE monitoring (separate effort)

Troubleshooting

Environment Recovery

When scripts fail due to missing modules, import errors, or a corrupt virtual environment, recover with:

cd "<powerpoint-skill-root>"
rm -rf .venv
uv sync

Resolve <powerpoint-skill-root> from the loaded skill location before running the command.

This recreates the virtual environment from scratch using pyproject.toml as the single source of truth. The Invoke-PptxPipeline.ps1 orchestrator runs uv sync automatically on each invocation unless -SkipVenvSetup is passed.

When uv itself is not available, install it first (see Installing uv above), then retry. When Python 3.11+ is not available, run uv python install 3.11 to have uv fetch and manage the interpreter.