PowerPoint Slides Skill
Turn academic papers or any structured content into visually rich .pptx presentations, driven end-to-end by your AI coding assistant.
The skill generates .pptx via local script execution (Node.js/PptxGenJS, Python, LaTeX, Graphviz), declares standard tools (Read/Write/Edit/Bash), and shows transparent data flow with no signs of exfiltration, credential access, or destructive defaults; however it runs arbitrary generated scripts via Bash without explicit least-privilege scoping, user-confirmation gates, or rollback guidance, and scripts are only partially shown, so full marks are not awarded.
Documentation preemptively covers many known pitfalls (PptxGenJS object mutation, '#' hex corruption, OMML blank in LibreOffice) and provides troubleshooting tables and error-recovery paths (manifest error fields), showing good self-consistency; but static review cannot confirm the scripts actually run, the dependency chain is very long (pandoc, xelatex, graphviz, mermaid-cli, soffice, sharp/react-icons) with no committed test suite — capped at 10 by static calibration.
Trigger design is exemplary: explicit positive/negative triggers including Chinese phrases ('做PPT', '做个报告'), exclusions for Beamer/TikZ, clear non-fit ranges, and a defined action vocabulary; deducted for environment fit — the toolchain relies on globally installed overseas packages with macOS/brew-oriented install notes and no guidance for mainland-China reachability or Windows.
Layered documentation is strong (SKILL.md workflow + references/ for themes/formulas/diagrams + scripts/), with progressive disclosure, troubleshooting, and an internal QA rubric; MIT license present; deducted for missing versioning, changelog, and explicit maintenance ownership/update path, and an unverified publisher.
The value proposition is clear (academic PPT with native OMML math, multi-engine diagrams, visual QA) and differentiated versus manual PPT authoring or Beamer; but static review cannot verify that key-path outputs are directly usable — quality hinges on a long multi-step script chain with no committed finished-output evidence, so capped at 7 and set to 6.
Auditable primary material is provided (detailed rules, script sources such as check_overlaps.py, manifest schemas, theme JSON) allowing partial cross-checking of claims against implementation; but there is no CI workflow, committed tests, or third-party execution evidence, and no independently reproducible output samples — capped at 5 by static calibration.
- Very long dependency chain (Node/PptxGenJS, pandoc, xelatex, Graphviz, Mermaid CLI, LibreOffice, Python lxml, etc.); each must be installed and verified before first run, and the brew-based instructions do not cover other platforms.
- OMML formulas render blank in LibreOffice previews by design — use XML validation for QA and always verify final rendering in real PowerPoint.
- The skill executes model-generated JS/Python via Bash; run in a controlled working directory and review generated scripts before execution.
- The core toolchain depends on overseas npm/LaTeX resources, which may be unreachable from mainland-China networks; no Windows support guidance is provided.
- No versioning or changelog exists; behavior may change on update — pin the repository revision you rely on.
What does this skill do, and when should you use it?
This is an AI coding assistant skill for generating visually rich PowerPoint presentations from papers, research notes, or any content. It is built on PptxGenJS (Node.js) with native OMML math, LaTeX-rendered image formulas, and a three-engine diagram pipeline (Graphviz/Mermaid/TikZ), plus 5 built-in color themes. The workflow spans a requirements interview, structure planning, formula and diagram preparation, script generation, and an automated PDF-to-image visual QA loop. With 29 hard rules and 5 mathematical slide templates, it targets academic talks, seminars, and defenses.
The skill reads the user's paper or material (Phase 0), runs a requirements interview via AskUserQuestion covering duration, audience, and theme (Phase 1), produces a structure plan that requires user approval (Phase 2), batch-renders formulas (OMML or 600 DPI PNG) via render_latex.py and diagrams via render_diagrams.py (Phases 3/3.5), generates and executes a full PptxGenJS script (Phase 4), then converts output via LibreOffice to PDF and pdftoppm images, dispatches parallel subagents for per-slide visual inspection, runs check_overlaps.py for boundary/overlap issues, and loops on a scoring rubric (Phase 5). It supports 9 actions including create/compile/review/audit/visual-check/validate/extract-figures; OMML placeholders are injected into native PowerPoint math by inject_omml.py via pandoc.
- A graduate student or researcher turning a paper into conference or lab-meeting slides quickly, with formulas intact
- A PhD candidate preparing a thesis defense needing reference slides, backup slides, and automated layout QA
- An instructor building math-heavy lecture slides who wants native, editable OMML formulas instead of blurry screenshots
- Frequent presenters who want slide quality enforced by fixed rules (density bounds, layout diversity, references slide)
- Users who want to batch-extract figures from a paper PDF and embed them with source attribution (optionally via pdf-mcp)
What are this skill's strengths and limitations?
- Mature formula pipeline: OMML native math stays editable after generation, with a 600 DPI LaTeX image fallback for complex paragraph formulas
- Five-layer diagram pipeline covers structural, behavioral, math, and extracted figures, with forced theme-color injection
- Complete quality system: 29 hard rules, scored QA loop, card-aware overlap/boundary checker scripts
- Thoughtful academic-specific design: references slide, backup slides, duration heuristics from 5 to 90 minutes, talk-type tips
- Ships with a complete real-world example (25 slides, 23 formulas) to gauge output quality
- Heavy dependency chain: Node.js, pandoc, a full TeX distribution, Graphviz, LibreOffice, and more
- Known limitation: LibreOffice does not fully render OMML, so native formulas appear blank/distorted in the QA preview
- SKILL.md relies on Claude Code-specific mechanisms (allowed-tools, AskUserQuestion, $ARGUMENTS, Agent subagents); porting to other assistants requires rework
- No automated test suite; adaptation to other AI coding assistants is mentioned but not documented
- Tuned for academic presentations — business/marketing slide styles are not a design goal
How do you install this skill?
Prerequisites: Node.js, pptxgenjs, pandoc, a TeX distribution, Graphviz, LibreOffice, and Python packages Pillow, lxml, markitdown[pptx]; optionally mermaid-cli, Ghostscript, and pdf-mcp. Install: git clone https://github.com/Noi1r/powerpoint-skill.git, then mkdir -p ~/.claude/skills && cp -r powerpoint-skill/powerpoint-slides ~/.claude/skills/, and restart Claude Code — the skill triggers automatically on powerpoint/pptx/PPT/slides keywords.
How do you use this skill?
Trigger it with natural language in Claude Code, e.g.: "Help me make a PPT based on this paper: /path/to/paper.pdf", "Make a presentation about X using the Midnight theme", "Compile generate_slides.js", "Review my slides: output.pptx", or "Extract figures from /path/to/paper.pdf pages 3-5". Trigger keywords include powerpoint, pptx, PPT, make a ppt, 做PPT, and related phrases.
How does this skill compare with similar options?
The skill explicitly positions itself against the Beamer/LaTeX slides workflow: its trigger words deliberately avoid beamer, .tex, and tikz (which belong to a separate beamer skill), and when a user just says "make slides" it asks whether they want PPT or Beamer rather than defaulting silently. If you need LaTeX-native slides, the author implies you should use the companion beamer skill.