Gabberflast/academic-pptx-skill: A Claude Skill That Replaces Design-First Slides With Argument-First Ones

The problem is that Claude writes topic labels, and academic talks need claims

Ask a general-purpose model for a conference deck and you tend to get a sequence of headings: Introduction, Methods, Results, Discussion. Each slide is defensible on its own and the deck as a whole says nothing. That is a reasonable default for a product update and a poor one for a twenty-minute talk where the audience needs to follow an argument and leave with a claim they can repeat.

This skill exists to flip that default. According to the README, when active it makes Claude write every slide title as a complete sentence stating the takeaway, what the project calls an action title, rather than a topic label. It also structures the deck as a logical argument following situation, complication, resolution, and applies what the README calls the ghost deck test: the action titles alone, read in sequence, should tell the full story.

The intended audience is narrow and identifiable. Conference speakers, seminar presenters, thesis defenders and people preparing grant briefings. If you are building a sales deck or a product launch, the constraints here (white backgrounds, three colours maximum, no decorative icons) will read as austerity rather than discipline.

How the skill works: routing logic plus three markdown layers

A Claude Skill is not a program. It is a folder of instructions that Claude loads when it decides the context matches. The repository layout reflects that. SKILL.md is described as the entry point holding routing logic and design standards. content_guidelines.md carries argument structure, action titles, citations and deck architecture. slide_patterns.md holds per-slide-type implementation patterns with PptxGenJS code. README.md is the file you are reading about.

The division of labour matters. The README states that this skill works alongside Anthropic's built-in PPTX skill, which handles the technical file generation, while this one handles content, argument structure and design standards. So the actual .pptx is produced by the other skill; this one shapes what gets written into it. If you disable or lack the built-in PPTX skill, you have a set of writing rules and no generator.

That layering is also where the sharpest limitation sits. The README does not document what happens when the two skills disagree, whether the academic constraints survive a long conversation, or how Claude resolves a user instruction that contradicts an action-title rule. Those are the questions a careful adopter would ask first, and the repository is silent on all of them.

Installing the academic-pptx-skill zip in claude.ai

There is no package manager here and no command line. The README gives a four-step manual install that starts with downloading the repository as a zip through the Code button and the Download ZIP option.

The remaining steps happen in the claude.ai interface. Navigate to Customize, then Skills, and upload the zip. After uploading, confirm the skill appears in your skills list and is toggled on. The README adds one requirement in a note: code execution and file creation must be enabled under Settings, Capabilities. Without that toggle, the companion PPTX generation has nothing to run in.

# No install command. Download the repository as a zip from GitHub
# (Code -> Download ZIP), then in claude.ai:
#   Customize -> Skills -> upload the zip
#   Settings -> Capabilities -> enable code execution and file creation

Usage after that is plain language. The README gives three example prompts, including "Make slides for my conference paper on X" and "Build a deck for my thesis defense". The README states that Claude detects the academic context, loads the skill automatically, and applies the guidelines before generating slides, and that you do not need to give any special instructions. What you should see is a deck whose titles are sentences rather than nouns. If the titles still read as topics, the skill did not load, and the first thing to check is the toggle in your skills list.

The citation and closing-slide rules are the parts most likely to bite

Two of the enforced standards have consequences beyond layout. The first is citation: the README says the skill applies in-text citations on every borrowed figure and a References slide at the end. That is the right default for academic work and an awkward one when your figures are your own, when a figure is under a licence you have not checked, or when the venue caps your slide count. The skill does not appear to distinguish between a figure you drew and one you reproduced, at least not in the README.

The second is the ending. The README states the deck should end on a Conclusions slide that stays on screen during Q&A, never on Thank You or a blank. This is a real convention in conference practice and a direct conflict with the habit many speakers have of ending on an acknowledgements slide or a QR code. If your institution or funder requires a specific closing slide, you will be overriding the skill by hand.

Neither of these is a defect in the skill. They are choices, and the README presents them as non-negotiable defaults rather than configurable options. There is no documented way to turn off the citation rule or substitute a different closing pattern, which means the skill is best suited to speakers who already agree with these conventions.

The PDF is the escape hatch, and it is the reason to look even without Claude

The repository ships Academic_Presentation_Best_Practices_DH_built_with_Claude_2_2026.pdf at the top level. The README describes it plainly: a PDF which summarizes the advice and does not require use of Claude.

That single sentence changes who the project is for. If your institution blocks third-party AI tools, if you are advising students who cannot use Claude, or if you simply want the rules without an account, the PDF is the deliverable and the skill files are the implementation. The README does not state whether the PDF and the markdown guidelines are kept in sync, and that is worth checking yourself before you cite one as the source for the other.

The background section names where the advice comes from: Barbara Minto's Pyramid Principle for structured argument and action titles, Naegle (2021), "Ten simple rules for effective presentation slides," in PLOS Computational Biology, and what the README calls standard consulting and academic presentation practice. Those are traceable sources rather than invented authority, which is more than many prompt collections manage.

Compared with Anthropic's built-in PPTX skill and a template-driven workflow

The obvious comparison is the skill this one is designed to sit beside. Anthropic's built-in PPTX skill handles technical file generation: it turns instructions into an actual presentation file. This skill handles content, argument structure and design standards. The difference in approach is that one is a renderer and the other is a set of editorial rules. You can use the built-in skill alone and get a file; you can read this repository's markdown and PDF alone and get no file at all.

A second comparison is a template-driven workflow, the kind where a lab maintains a .potx or a LaTeX beamer theme and everyone fills in the blanks. That approach guarantees visual consistency and enforces nothing about argument. A title placeholder labelled Results will accept a topic label or a claim with equal indifference. This skill attacks the opposite problem: it constrains what the title says and leaves the visual system deliberately thin, with the README specifying white backgrounds, a single sans-serif font, three colours maximum and no decorative icons.

Neither approach is strictly better. Templates scale across a group without any AI account. This skill scales across topics for one author who is already working inside Claude.

Maintenance, licence and what the repository does not tell you

The repository is not archived and the last push was on 2026-07-14, which is recent enough that the instruction files are plausibly still aligned with current Claude behaviour. There are no retrieved releases, so upgrades are whatever lands on main. That is a real cost: skills are instructions, and a change to the action-title rule or the citation rule arrives without a version number to pin. If you adapt the files for your own lab, keep your edits in a fork so a pull from main does not silently overwrite them.

The licence is MIT, stated in both the README and the repository. MIT permits use, adaptation and sharing, and it comes with no warranty. Nothing here is legal advice, and one practical point deserves a check rather than an assumption: if you redistribute a deck built with this skill, the licence covers the skill's files, not the figures you placed in the slides or the fonts and images you used. Those carry their own terms.

What the documentation does not cover is the operational side. There is no changelog, no compatibility statement naming which Claude surfaces support custom skills, and no guidance on what happens when a deck spans more slides than a single generation can hold.

Editorial conclusion

Adopt this if you already write in Claude and your problem is that generated decks read like topic lists rather than arguments. Skip it if you need a build pipeline, a template system, or anything that runs outside claude.ai, because the repository ships markdown guidance and a PDF, not code. Before trusting it, open content_guidelines.md and confirm the action-title rule matches how your field writes claims, then run one real deck through it and read the resulting titles in sequence.

Frequently asked questions

Official sources

1. Gabberflast/academic-pptx-skill on GitHub
2. Issues
3. License: MIT
4. README