zhouwei713/luobo-ppt

What it solves

AI-generated presentations often come out as flat images (impossible to edit), or look template-generic, or shrink text until it's unreadable, or silently rewrite your original wording, or lose formulas and numbered notes. This project turns those habits into checkable rules before you get the file.

How it works

Two routes share one design engine. In the beautify route, it reads an existing .pptx and extracts every bit of content into a structured deck.json (per-page content) plus a style.json (colors, fonts, layout). In the generate route, it takes a topic, article, or script and writes the deck itself. Both routes go into plan.json (layout planning), then the scripts draw illustrations, assemble a native PowerPoint file, render a preview, run automatic checks, and give a five-dimension self-review.

Scripts handle the deterministic work (layout, fidelity checks, rendering), while the Agent handles content understanding and aesthetic judgment. The project guarantees that not a single character from the original is lost on beautify: it stores every paragraph and compares page-by-page after finishing, raising an error naming the page and sentence if anything is missing. It also enforces per-layout capacity limits (e.g., a funnel maxes at 5 levels, each label under 26 characters) during outlining, so content that doesn't fit is caught early.

Rendering previews defaults to LibreOffice; if that's not installed, it draws a close approximation from layout data (12 pages in about 10 seconds). It can run without any image-generation model—backgrounds, decorations, and illustrations are drawn programmatically—and can optionally call OpenAI, Gemini, Volcano Ark, HiAPI, or any OpenAI-compatible endpoint for richer illustrations, falling back to programmatic art if a single call fails. For photo needs, it can search public image libraries (CC0, public domain, CC BY, CC BY-SA, plus Pexels/Unsplash/Pixabay with a key), crop photos to fit, and list credits. It also supports animated reveals for scripted or teaching decks (cards and steps appear one at a time), with a fade transition between slides; static previews can be supplemented with storyboard GIFs.

The output is always a native, editable .pptx: text is real text boxes, tables are real tables, charts are PowerPoint-native charts you can double-click to edit. It works in PowerPoint, WPS, and Keynote.

Who it's for

This is for anyone who wants to create or restyle presentations through an AI coding agent (Claude Code, Codex, Gemini CLI, Cursor, Copilot, or any agent that supports SKILL.md). It suits talk-track videos, articles turned into talking decks, plain work reports or courseware that need a style overhaul, science/math slide decks with many formulas, AI-tool tutorials with lots of screenshots, data reviews and competitor analyses needing native tables/charts, and conference-room projections that need larger minimum font sizes.

Highlights

• Two working modes: generate a deck from a topic or article, or beautify an existing .pptx following a reference image—original text stays untouched.
• Always produces native editable .pptx: real text, real tables, double-clickable PowerPoint charts.
• 14 built-in visual styles (mint-fresh, business-blue, warm-orange, dark-tech, lavender-soft, paper-notes, black-gold, geo-blocks, cobalt-type, ink-lime, pastel-mist, wave-launch, ink-magazine, swiss-grid) that you can pick by name.
• Strict fidelity QA: no lost characters, no overflow, no contrast problems; a self-review scores uniformity, hierarchy, detail, communication, and rhythm with concrete evidence.
• Programmatic illustration — works without any AI image model, and can optionally use commercial-safe image search or your own photos.
• Formula preservation: converts formula objects into plain-text math so they display and edit in any software.
• Special layouts for screenshots (with numbered callouts) and for rich content like timelines, funnels, hub-and-spoke, 2x2 matrices, SWOT, PEST, Porter's Five Forces, and business model canvas.
• Supports company-brand templates (.pptx or .potx): keeps logos, color bars, footers from the master, uses the template's theme colors and fonts, and auto-adjusts text color for dark or bright covers.
• Optional "quick mode" to skip intermediate confirmation steps; extensive update log shows continuous hardening (e.g., handling wide numbers, bright brand colors, hand-drawn lines, per-page locking of manual edits, and template reading).