How scholar-slides Uses HTML to Build Accurate Academic PPTs

Author Lu Gong explains the motivation behind creating scholar-slides: after years of struggling with AI‑generated PPTs that sacrifice academic fidelity for visual flair, he sought a skill that meets the priorities of scholarly presentations—high fidelity, evidence‑based content, and editability.

He evaluated six open‑source slide skills (baoyu‑slide‑deck, guizang‑ppt, huashu‑design, ppt‑master, frontend‑slides, and his own paper‑slide‑deck) and found each lacking in essential academic features such as proper formula rendering, accurate figure extraction, reliable citation formatting, and systematic quality control.

His earlier paper‑slide‑deck mixed raw screenshots with generated images, leading to failures when formulas or numbers were rendered as pixels, tables appeared as images, and citations were often incorrect.

scholar‑slides replaces that approach with an HTML‑first pipeline. Content is first written in HTML, then exported using a stack of reveal.js and KaTeX. Formulas become true vector text selectable in PDFs; tables are built from extracted data using native <table> tags; figures are cropped from the original PDF based on precise bounding‑box coordinates, preserving figure numbers and sources; references are resolved via Zotero with Crossref fallback, producing verified BibTeX entries.

Export is handled by Playwright, which prints each HTML slide to a vector PDF page and also generates a PPTX file where text, tables, and notes remain editable, while formulas and images are embedded as pictures. Each export runs a parity check to ensure no content loss, and the tool supports bilingual (Chinese/English) decks.

The HTML route enables deterministic code‑based quality checks. A static validator blocks unregistered layouts and fake SVG text. Playwright measures page geometry, confirming that empty space and minimum font size (12 px) meet defined thresholds—something impossible with a pure image‑based workflow.

Concrete example: In a GLM‑5 deck, four slides had unreadable figures. Initial intuition blamed figure width, but Playwright revealed that fixed elements (title, caption, footnote) compressed a 936 px vertical canvas to only 450 px for the figure. By enabling a “hero” mode that folds captions into the footer, about 200 px were reclaimed, raising the smallest text from 8.7 px to 12 px.

scholar‑slides enforces an integrity gate: numbers, citations, and figures must be verified; missing or unverified items are marked with clear placeholders (MISSING or UNVERIFIED). The pipeline consists of seven stages with three manual checkpoints—checking key points, narrative flow, and a final authenticity signature—so all doubts are presented for human review.

Two design tweaks improve usability: each slide title must be a conclusion sentence, creating a logical chain when read sequentially; and the tool auto‑generates speaker notes with bilingual speech‑duration estimates, helping presenters align with time budgets.

In a benchmark of ten papers with varied layouts, scholar‑slides showed zero crashes, a 98 % success rate in locating figures, and roughly 92 % of cropped figures being directly usable. Single‑column and arXiv layouts achieved >95 % success, while IEEE double‑column dense pages reached ~75 % success, with low‑confidence figures flagged at the first checkpoint.

The skill is available at

https://github.com/luwill/research-skills/tree/main/scholar-slides

. After cloning, run ./install.sh to install. Users can invoke it in Claude Code, Codex, or similar agents by providing a PDF or arXiv link and requesting a group‑meeting PPT; the process pauses at three verification points for user confirmation.