Best document-to-slide-deck skills for AI agents

An agent asked to turn a document into a deck makes one decision before any of the visible ones, and makes it silently: what happens to the document. Left to its own general knowledge it does the obvious thing — reads your paper as a topic, writes fresh bullets that sound about right, and hands back slides. The results page now carries approximately your numbers. The architecture diagram is a redrawing of your diagram. Nothing looks wrong until a co-author opens the file to fix a typo and finds that the slide is a single flat image.

So shop for a deck skill by how much of your source arrives on the other side — your words, your figures, your editability, and, as it turns out, your name. One line found by reading these instructions is worth stating up front: two skills that share a name each stamp a different company onto your title slide unless you tell them otherwise.

Top picks

Six picks, plus the alternatives and same-name copies discussed below, with publisher, licence and last update side by side.

skill	publisher	license	verdict	updated
ppt-creator	daymade	MIT	Top pick — editable objects	2026-07-27
aris-paper-slides	OpenLAIR	NOASSERTION	Top pick — editable objects	2026-07-16
slides-creator	daymade	MIT	Top pick — generated images	2026-07-27
scientific-slides	K-Dense-AI	MIT	Top pick — generated images	2026-07-28
codex-ppt	ningzimu	MIT	Top pick — generated images	2026-07-23
image-to-editable-ppt	ningzimu	MIT	Top pick — images back to objects	2026-07-23
baoyu-article-illustrator	JimLiu	MIT	For articles, not decks	2026-07-04
academic-slides	EvoScientist	Apache-2.0	Alternative	2026-07-26
baoyu-slide-deck	JimLiu	MIT	Renders slides-creator's output	2026-07-04
Journal Club Ppt	xuzhougeng	AGPL-3.0	Check first	2026-07-28
scientific-slides	synthetic-sciences	Apache-2.0	Check first	2026-07-27
scientific-slides	foryourhealth111-pixel	Apache-2.0	Check first	2026-07-22
Scientific Slides	jaechang-hits	NOASSERTION	Check first	2026-07-24

ppt-creator: the one that leaves you something to edit

ppt-creator builds a markdown slide file, a separate notes file, a citations file and chart images, then exports PPTX where python-pptx is available — text objects and images, not a picture of a deck. Headings must be testable assertion sentences rather than topic labels, speaker notes are written to a 45-60 second spoken length, and the deck is scored against a bundled rubric with automatic refinement below 75.

Its honesty rule is the reason it sits on this list: when the data you supplied is not enough for a chart, it is instructed to emit a placeholder plus a "list of missing fields" rather than a plausible-looking graph. The caveat is the entry point — it will happily start from a bare topic and safe defaults, so the discipline protects the chart, not the argument.

aris-paper-slides: for a conference talk you can still edit the morning of

aris-paper-slides is the pick that hands back two artefacts and keeps both of them real: a Beamer deck compiled to PDF, and an editable PowerPoint built with python-pptx that carries the speaker notes across into the notes pane. Your figures are not redrawn — the paper's own figure files are linked into the slide directory and included by reference — and its rules forbid fabricating data, requiring every number on a slide to come from the paper's own section sources.

It budgets the talk rather than the deck: talk format and length are settings, slide count follows from them, and alongside the notes it writes a word-for-word script with per-slide timings and an anticipated question list. Two costs. It expects a compiled LaTeX paper on disk, so it is no help if your source is a Word file, and it stops for confirmation at the outline before drafting anything. And resolve the licence yourself before you build on it: the file declares MIT while the repository record carries no assertion at all.

slides-creator: for keeping your own words, if you can live with image slides

slides-creator opens with a rule it calls the First Law, and the first step under it is to collect the user's original words — transcripts, articles, notes, previous decks — before any structure work happens. Its stated weight order runs from the user's own words down through approved external material and AI synthesis to AI invention, and it names the product of skipping that order "polished, plausible, unusable garbage". The premise is blunt: "AI cannot write high-quality content for the user. It can only help the user express their own content better." If you have nothing written down it does not refuse the job — it says the whole narrative then has to come out of your head through structured conversation. What it will not do is fill the gap itself.

The trade-off is the artefact. Rendering is delegated entirely to baoyu-slide-deck, which generates "professional slide deck images" designed for reading and sharing "rather than live presentation", and the PPTX you get back is those images merged. That puts this one on the picture side of the choice below, with the usual consequence: its own failure log pairs "Text unreadable in images" with the root cause "Model doesn't support Chinese well", and the prescribed cure is a corrected prompt, not an edit. You also get an extracted speaker-notes file; plan on it carrying the talk.

scientific-slides: good for a research talk, if you take the images and change the author line

Among the copies circulating under that name, K-Dense-AI's scientific-slides is the one that declares a version — 1.3 — and still keeps both a generated-PDF route and a PowerPoint route. Its results-slide instructions tell the agent to list the working directory for existing figures and attach the real chart before generating anything. That is a genuine fidelity step: your figure enters the prompt instead of being described from memory.

Know which route you are on, though. The one it recommends is to "Generate each slide as a complete image using Nano Banana Pro, then combine into a PDF"; the object route is filed underneath as the alternative for when "you need editable slides (e.g., for company templates)". On the recommended path your attached chart still passes through the image model, so this belongs on the picture side too, and a typo means a regenerated slide.

And the default byline is the skill author's: the file instructs Default author is "K-Dense" unless another name is specified. Change it, or present someone else's company on your title slide.

codex-ppt: when the deck is meant to be looked at

codex-ppt makes every slide a complete generated 16:9 image and assembles them into PPTX, and it is unusually clear about the consequence: "Do not use it when every textbox, chart, or shape must remain separately editable." That sentence belongs in more skill files than it appears in.

What earns it a pick is how it treats your material inside a generative pipeline. Supplied figures, screenshots and logos are handled by a dedicated step that says to "Treat required assets as strict inputs" and confirm slide-to-asset mapping before anything is generated. Local drawing, HTML screenshots and python-pptx layouts are listed as failure modes rather than fallbacks, so a missing backend produces a reported blocker instead of a quietly degraded deck. Expect approval gates: outline, style, backend, and one sample slide.

image-to-editable-ppt: for the deck you already have as pixels

image-to-editable-ppt is the repair path, and it is the reason the raster-versus-objects choice is worth thinking about before you make it rather than after. It takes images, PDFs or an image-based PPTX and rebuilds them into object-level PowerPoint through an editppt CLI, decomposing each page, reconstructing it, and validating the result against a manifest that must independently rebuild the page.

The detail that matters is what it refuses. The obvious cheat — drop the original screenshot in as the background and float text boxes over it — is explicitly ruled out: text overlaid on a full-slide source raster "is not an acceptable fallback", and the deck-level check hunts for exactly that pattern. Cost of entry: the CLI, an OCR token for good text fidelity, and subagents for multi-page input.

A skill called scientific-slides can be any of these

Skill names are community-published and not namespaced, so a name tells you the topic and nothing about the document. Ask an agent for scientific-slides and you might get a design-and-delivery reference with Beamer templates and no image model in it anywhere, an experimental cut with the PowerPoint and Beamer routes removed, or either of two copies of the same generation pipeline at different stages of maintenance.

Here is the check that pays for itself. K-Dense-AI's copy instructs the agent that the default author is "K-Dense". synthetic-sciences' copy, which declares nothing at all about where it came from, carries the identical instruction with its own company in the slot: Default author is "Synthetic Sciences". Neither line sits in a legal or attribution section; both are items in a numbered list about formatting consistency, several screens down. Nothing in either file's name, description or front matter warns you, and the first time you find out is when someone asks who that company on your title slide is.

The rest of the check takes about a minute: foryourhealth111-pixel's copy declares the same author as K-Dense-AI's and no version, and is the larger file while still carrying inline the five sections the versioned copy has moved out into reference files, so the bigger document is the older one; jaechang-hits' copy ships Beamer templates and design references instead of a generation pipeline; and more than one of them declares a licence its repository does not.

What to check before you hand a skill your document

Four questions, each answerable in under a minute from the skill's own text.

Does it require your material, or accept a topic? A skill that begins by asking for your transcripts and drafts will produce something you recognise. A skill instructed to "gather context or apply safe defaults" will produce something confident. Both are legitimate; they are not interchangeable.

Are the slides objects or a picture? Look for python-pptx, Beamer, Marp or a markdown slide file on one side, and image generation plus assembly on the other. If it is images, accept that typo fixes mean regeneration — and check that the skill accepts it too. baoyu-article-illustrator puts it most plainly, because illustrating an article is the same problem one image at a time: "Never repair rendered text by painting over a generated bitmap." A garbled label means a new prompt file and a new output path, with the flawed candidate kept for comparison — not an ImageMagick patch that looks fine until someone zooms in. It is worth installing on its own account if you write prose rather than decks; it reads the article first, decides where a visual earns its place, writes each prompt to disk before any backend is called, and requires the article's "actual numbers, terms, metrics, quotes" to travel into the prompt.

Whose figures end up on the results slide? Journal Club Ppt lifts figure panels straight out of your PDF rather than regenerating them, and organises the deck around the paper's own line of argument — at the cost of an AGPL-3.0 licence you need to be comfortable with. aris-paper-slides includes your paper's figure files directly as well. The other picks above put your chart through an image model, K-Dense-AI's included on its recommended route.

Whose name is on it? Default bylines, default authors and default watermark settings are set in the skill file, not in your request.

One more, for research talks: subtraction is a feature. academic-slides instructs the agent to "Cut 80% of your paper's content" and treats a deck you could rebuild the paper from as a deck carrying too much. Fidelity to a source is not the same as carrying all of it across.

Getting your document to the other side

The silent decision at the start of this is now yours instead of the agent's. If the document is the point — a paper, a transcript, a set of numbers you can be held to — the two that hand your material back as things you can still edit are ppt-creator, for a deck built out of a markdown slide file and exported where python-pptx is available, and aris-paper-slides, for a talk compiled from a LaTeX paper with your own figure files in place. Reach for slides-creator when what needs protecting is your voice rather than your file — it will not write your content for you, and what it returns is a deck of images — and K-Dense-AI's scientific-slides for a research talk on the same understanding, with the author line changed. If the deck exists to be looked at rather than edited, codex-ppt will tell you so itself. And if you already have a deck made of pixels, image-to-editable-ppt is the way back.

What you have avoided: a beautiful deck of approximately-your-numbers, uneditable, presented by a company you have never worked for.