Convert to pptx

Skill from2001/useful-skills/.claude/skills/convert-to-pptx

Convert a slide image (PNG/JPEG) or PDF into an editable PowerPoint (.pptx) that visually matches the source — rebuilt from native shapes and text boxes, then render-compared against the original. Use when the user asks to turn a slide image, screenshot, or PDF into editable PowerPoint, or to 'rebuild this as a slide'.From its SKILL.md

Install

npx -y skills add from2001/useful-skills --skill convert-to-pptx

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.6 KB, as published. Nobody here has run it

Convert to PPTX

Reproduce a slide image — or every page of a PDF — as an editable PowerPoint deck.

The two non-negotiables

Editable output. Rebuild the slide from native PowerPoint primitives — text boxes for text, shapes for frames/bands/cards/backgrounds, cropped PNG/JPEG only for icons and illustrations. No SVG. Never paste the whole slide as a single image. If the user can't click a title and retype it, the skill has failed.
Render and compare before delivering. PowerPoint layout differs from your mental model — fonts substitute, text reflows, positions drift. After building the deck, render it back to images, eyeball-compare against the source, fix what looks wrong, and re-render. At least one fix-and-rerender cycle. Don't ship a first pass.

Inputs

Image (.png, .jpg, .jpeg, .webp) → 1-slide deck
PDF (.pdf) → N-slide deck, one slide per page (works even for image-only PDFs)

Tooling

Python: python-pptx (build), Pillow (crop icons), pymupdf (PDF → PNG, only if input is PDF)
LibreOffice (soffice) headless for the render-back step. Install with brew install --cask libreoffice / apt install libreoffice / winget install TheDocumentFoundation.LibreOffice. If it's genuinely unavailable, fall back to unoconv or PowerPoint's CLI — but do not skip the verification step.

Sandbox note. LibreOffice writes profile data outside the project directory (e.g. ~/Library/Application Support/LibreOffice on macOS, ~/.config/libreoffice on Linux), so render_pptx.py will fail when Claude Code's sandbox is enabled. If the render step errors out with a permission/sandbox failure, ask the user to approve running render_pptx.py outside the sandbox (or to disable the sandbox for this turn). Do not silently skip the verification cycle.

Helpers in scripts/:

pdf_to_images.py input.pdf <outdir>/ --dpi 200 — split a PDF into per-page PNGs
render_pptx.py output.pptx <outdir>/ --dpi 150 — render the produced deck back to PNGs
verify_pptx.py output.pptx — assert no SVG, no full-slide image, real text frames, multiple shapes per slide

Workflow

For a PDF input, split it with pdf_to_images.py first, then loop the per-slide workflow over each page, accumulating slides into one Presentation() object (don't build N decks and concatenate).

Per slide:

Decompose the source. Look at the image and list the elements that actually exist — background, title, message/callout boxes, cards, headings, big numbers, bullet lists, bands, alerts, conclusion box, page number, icons. Note rough position and size for each.
Classify each element. Text → editable text box. Frames/bands/cards/backgrounds → shapes. Icons/illustrations → cropped PNG/JPEG from the source (use Pillow). If an icon is too small or blurry to crop cleanly, generate a replacement with the nano-banana skill or any image generator and treat it the same way.
Build the slide. Set slide size to 16:9 (13.333" × 7.5" — python-pptx defaults to 4:3, so set this explicitly). Place background, then shapes, then text boxes, then cropped icons. Sample real colors from the source image rather than guessing.
Get the text exactly right. No typos. Preserve full-width vs half-width characters, numbers, units, and punctuation as they appear. Size the text box so headings don't unintentionally wrap.
Save the deck next to the input file (report.pdf → report.pptx).
Render and compare. Run render_pptx.py, view the rendered PNGs alongside the source, and look for anything visibly off — title placement, card sizes, big-number scale, line spacing, band height, unintended text wrapping, margins. Fix what's wrong, re-save, re-render. At least one cycle.
Final check. Run verify_pptx.py — it'll fail the deck if SVG sneaked in, if a full-slide image is present, if text got flattened, or if the slide has only one shape (the screenshot trap).

For long PDFs (20+ pages), ask the user whether they want every page rebuilt or only a subset before starting.

Deliverable

An editable .pptx next to the input, no SVG, no whole-slide image, all text editable, every shape individually selectable, 16:9, with no obvious visual discrepancy from the source after at least one fix-and-rerender pass.