In Practice: Make a slide deck that looks like yours

AI for Research Efficiency, In Practice, with Douglas Hutchings. Narration: an AI-generated voice (ElevenLabs).

One real run, replayed: an agent asked for the best way to make a deck from a paper that looks and sounds like its presenter, before building anything. Three helpers looked into the options, a reviewer tested them against the real files, the presenter decided, and the agent wrote the process down as a skill with a style brief in it. Then it built an editable PowerPoint deck in the presenter's fonts, colors and voice, with a chart that builds in steps, and a script checked the deck against the brief. Recorded on 28 October 2026 with Codex 0.164.0 on the studio's Linux server; the same steps in Claude Code and Antigravity are below, from their own documentation (only Codex was recorded). The presenter's answers to the agent's questions were written before the run, from the studio's defaults.

1. The practice folder

Download it here: deck-practice.zip. It holds one open paper and its data, and the style of the person presenting it:

deck-practice/  AGENTS.md          the instruction file (Codex and Antigravity read AGENTS.md; for Claude Code, copy it to CLAUDE.md)  paper/             Piwowar et al. (2018), "The state of OA", PeerJ (run fetch.py to download it), CC BY 4.0  data/              its data: 100,000 articles with Crossref DOIs and their open-access status (fetch.py), CC0  brand/             the presenter's website style, the look of the series, and the font with its license  voice/             a sample of the presenter's own writing  second-field/      a second author, made up for this film, and EPA's table of methane from rice fields  MANIFEST.csv       every file's title, authors, year, license, address and checksum  fetch.py           downloads the paper and the data, and checks them

Two things in the data trip up a first read, as they did in the run: its lines end the way old Macs ended them (a carriage return, no line feed), and its text is not UTF-8.

2. Set up (Windows first, the Mac beside it)

You need Python (episode 2.0 installs it) and python-pptx (MIT; it makes PowerPoint files without PowerPoint).

Windows (PowerShell), in the practice folder:py -m venv .venv.venv\Scripts\python -m pip install python-pptxpy fetch.py

macOS (Terminal), in the practice folder:python3 -m venv .venv.venv/bin/python -m pip install python-pptxpython3 fetch.py

To look at the slides as pictures, the agent needs something that can draw them:

• PowerPoint itself, if you have it: it draws your fonts as your audience will see them. File > Export saves every slide as a picture; an agent can do the same from a script (Presentation.SaveAs, with 18 for PNG).
• LibreOffice, free: on Windows winget install TheDocumentFoundation.LibreOffice (its package installs for the whole machine, so Windows may ask for an administrator); on the Mac, the installer from libreoffice.org (not yet tried by the studio). It swaps any font it does not have, so install your fonts first (their static files, not only the variable one), and judge type in PowerPoint.

3. Ask for the process first

The request, typed in plan mode, word for word:

What is the best way to make a deck from this paper that looks and sounds like me? Interview me first, and don't build anything yet.

Then the run's next requests, as typed:

Use three subagents, one for each question: what the vendors themselves do so that decks don't all look alike; the ways to make a PowerPoint file that stays editable, with a chart from the data; and how a chart can move from step to step in a deck that stays editable. Each brings back its options with their trade-offs and sources. Wait for all three, then summarize.

Now start one reviewer subagent. Give it the options and this folder. It should test each option against the real files, not against what it expects: the fonts and colors in brand/, my writing in voice/, the data in data/ and the paper. It reports only the problems that would change the result, and how it found each.

Then the presenter's decisions and the skill, in one message, out of plan mode (Shift+Tab again):

My decisions: the series' font, Atkinson Hyperlegible Next, for everything; the bright orange as a tint only; speaker notes in my own voice, in the first person; about eight slides. For the other problems the reviewer found, take your recommended fixes. Write the process we agreed as a skill in .agents/skills/deck-in-my-style/SKILL.md, with a name and a quoted description at the top, and the style brief in it: the sources in the order that wins, the fonts and their licenses, each color's contrast and its job, how full each kind of slide may be, the rules for figures, titles as sentences, my writing sample for the voice, and a dated list of looks to avoid with their sources. Don't build the deck yet.

The build, checked at the first step:

Now make the deck with the skill: the paper's main findings in about eight slides, Figure 2 as a chart that builds in steps with Morph, then the same data as an editable chart. Show me the first slide rendered before you do the rest.

In the run, the agent asked to install python-pptx in a virtual environment inside the folder (approved), showed the first slide, and got "Looks right. Go ahead with the rest." Then /permissions > "Approve for me", which Codex describes on the screen as "Only ask for actions detected as potentially unsafe" (its documentation calls the preset Auto). Fresh eyes on every slide, then the check:

Render every slide to a picture and have a subagent that hasn't seen your code compare them with the brief and list every problem it sees. Fix what it finds and render again.

Check the deck against the brief, and tell me what you could not check.

The run took about 56 minutes from the first request to the checked deck.

Before the run, the studio listed nine traps it knew were in this folder (a disagreement between the presenter's website and the series on fonts, a brand color under 3:1, the data's line endings and encoding, the paper's own counting rules, charts that don't morph, python-pptx's blank deck, a preview that swaps fonts, the paper's credit), and kept the list out of the run's folder. The reviewer found three of them, and part of a fourth (the paper's years, not its rule about a test publisher's rows). The other five came up at other steps: the fonts in the agent's first look, the orange while it wrote the brief, the charts in a helper's report, python-pptx's blank deck in the build, the credit in the plan. The reviewer also found a problem the list had missed: the paper's own Figure 2B stacks the closed articles too, up to 100%. The test publisher's rows got past every step, and past the check (below).

4. The skill: deck-in-my-style

Download the folder: deck-in-my-style.zip. Put one copy in .agents/skills/deck-in-my-style/ (Codex and Antigravity) and another in .claude/skills/deck-in-my-style/ (Claude Code); on Windows, a copy, not a link.

deck-in-my-style/  SKILL.md                      the process: brief first, titles first, build, render and look fresh, check, hand back  references/brief-template.md  the style brief to fill in with your agent (its json block feeds the check)  references/banned-defaults.md looks that read as machine-made, dated 27 October 2026, each with its source  scripts/check_deck.py         the check (Python's standard library only)  scripts/pptx_helpers.py       what python-pptx has no setting for: theme, 16:9, titles, alt text, Morph, chart box, years

The brief puts your sources in the order that wins, so a disagreement has an answer. In the run, the presenter's website sets Lato and Open Sans and the series sets Atkinson Hyperlegible Next; a 2025 list of fonts to avoid names both of the website's. The brief wins over the list either way: Anthropic's own design skill says "the brief's own words always win, including when it asks for one of these looks".

5. A chart that builds in steps

PowerPoint's Morph transition moves the same objects from one slide to the next; "charts don't morph, they cross-fade" (Microsoft). So the chart that builds is drawn as shapes:

• each bar a rectangle with a name that starts with two exclamation marks (!!bar-2015), the same on every step;
• one step per slide, each slide with its own title, and Morph on every step after the first;
• then the same numbers as a real PowerPoint chart on the next slide, which you can still edit.

python-pptx has no setting for transitions; pptx_helpers.py writes Morph the way PowerPoint does, with a fade for older versions. Morph plays in Microsoft 365, PowerPoint 2019 and 2021, and PowerPoint for the web (for files on OneDrive for work or school or SharePoint); PowerPoint 2016 without its updates "plays a Fade transition".

6. Check it against your brief

Windows:  .venv\Scripts\python .agents\skills\deck-in-my-style\scripts\check_deck.py deck\talk.pptx --brief brief.md --sample voice\sample.md --out check.mdmacOS:    .venv/bin/python .agents/skills/deck-in-my-style/scripts/check_deck.py deck/talk.pptx --brief brief.md --sample voice/sample.md --out check.md

On the run's deck, the script found 37 failures: 24 text runs under 18 pt (23 of them source and credit lines), 8 titles a screen reader can't find (each title is a text box, not the slide's title placeholder), 4 pale chart fills under 3:1 against the background, and 1 chart without alternative text. What it can't see: the data (the chart keeps 141 rows from a test publisher that the paper's code leaves out, so 2000 reads 23.56% where the paper's method gives 24.24%), and the ink lines the deck draws between the pale fills. The same script on the before and after (two fresh sessions, the same one-line request, the second with run A's skill): 203 failures on the deck made without a brief, 41 on the one made with it.

Then two looks only you can give: PowerPoint's Accessibility Checker (Review > Check Accessibility), and a last pass in PowerPoint itself. The studio has no PowerPoint on its server: every slide picture in the film is a LibreOffice render, and the film does not claim that Morph played or what the Accessibility Checker says. The presenter's own check on Windows (Morph, PowerPoint's export, its Accessibility Checker, bold titles) is still to come. If your deck will be posted by a public university, the Justice Department's ADA Title II rule applies to what is posted after the compliance date (April 26, 2027 for entities of 50,000 people or more): its own example, a posted PowerPoint, "would probably have to meet WCAG 2.1, Level AA". Whether that reaches a lab's deck is for your campus to say. (Read 24 October 2026; the dates are being argued in court.)

7. Fonts and embedding

• An open font (SIL OFL) may travel inside the deck: "So can I embed OFL fonts in my document? Yes, either in full or a subset." In PowerPoint for Windows: File > Options > Save > Embed fonts in the file, and choose "Embed all characters (best for editing by other people)". PowerPoint for Microsoft 365 for Mac, 2024 and 2021 for Mac can embed too.
• A font marked Restricted cannot be embedded; one marked Preview/Print opens the deck read-only elsewhere.
• Google Slides "ignores fonts embedded in an uploaded PowerPoint file".

8. Same process, another field

The same skill with a fictional author's brief (an extension agronomist, made up for this film: Lato, field colors, short sentences in the second person) on EPA's table of methane from rice cultivation by state, 1990 to 2020. The EPA table is a US Government work; EPA's terms allow non-commercial, scientific and educational use.

In the run (a fresh Codex session, the same day), the agent noticed whose brief the skill held: "The skill contains a previous presenter's brief, so I'll use this folder's Lato, field colors, and plain grower-facing language." It wrote the PowerPoint file with Python's standard library alone (no python-pptx), six slides with two editable charts, and the same check found 21 failures (12 text runs under 18 pt, 6 titles outside the title placeholder, two charts and a table without alternative text).

9. The line not to cross

• Only fonts, pictures and logos you may use. A university's logo is usually for its authorized units: the University of Arkansas, for one, says "Students and student groups are not allowed to use these elements without special permission from university relations."
• Never let a number on a slide drift from its source, and keep the qualifiers.
• If your venue or your institution asks you to say where AI helped, say so.

The vendors' own slide tools

Anthropic's pptx skill (source-available, not open source; it builds decks with the Node library pptxgenjs; in Claude Code, /plugin install document-skills@anthropic-agent-skills). OpenAI's Presentations, in the ChatGPT desktop app (@Presentations). Gemini in Google Slides makes native Google Slides decks and takes a reference deck. Copilot in PowerPoint uses a Brand Kit and your template.

Sources

Read on 24 October 2026 unless dated otherwise; the tools' and vendors' pages read again on 28 October 2026.

• Piwowar H, Priem J, Larivière V, Alperin JP, Matthias L, Norlander B, Farley A, West J, Haustein S (2018). The state of OA: a large-scale analysis of the prevalence and impact of Open Access articles. PeerJ 6:e4375. https://doi.org/10.7717/peerj.4375 (opens in a new tab) (CC BY 4.0). Data: https://doi.org/10.5281/zenodo.837902 (opens in a new tab) (CC0); code: https://github.com/ourresearch/oadoi-paper1 (opens in a new tab) (MIT). Read 27 October 2026.
• Anthropic, "Improving frontend design through Skills" (2025): https://claude.com/resources/articles/improving-frontend-design-through-skills (opens in a new tab); the frontend-design and pptx skills: https://github.com/anthropics/skills (opens in a new tab).
• Microsoft: Morph (https://support.microsoft.com/en-us/powerpoint/training/use-the-morph-transition-in-powerpoint (opens in a new tab), tips and tricks https://support.microsoft.com/en-us/powerpoint/morph-transition-tips-and-tricks (opens in a new tab)); accessible PowerPoint presentations (https://support.microsoft.com/en-us/accessibility/powerpoint/make-your-powerpoint-presentations-accessible-to-people-with-disabilities (opens in a new tab)); embedding fonts (https://support.microsoft.com/en-us/office/benefits-of-embedding-custom-fonts-cb3982aa-ea76-4323-b008-86670f222dbc (opens in a new tab)).
• W3C, WCAG 2.2: https://www.w3.org/TR/WCAG22/ (opens in a new tab). ADA.gov, the Title II web rule: https://www.ada.gov/resources/2024-03-08-web-rule/ (opens in a new tab).
• python-pptx: https://python-pptx.readthedocs.io/ (opens in a new tab). Marp CLI: https://github.com/marp-team/marp-cli (opens in a new tab). Slidev: https://sli.dev/guide/exporting (opens in a new tab). Pandoc: https://pandoc.org/MANUAL.html (opens in a new tab). Quarto: https://quarto.org/docs/presentations/powerpoint.html (opens in a new tab).
• Google (Gemini in Google Slides, 30 June 2026), OpenAI (Create or revise a slide deck), Microsoft (Copilot in PowerPoint): the addresses are in the film's research/sources.json.
• Wikipedia, "Signs of AI writing": https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (opens in a new tab). Michael Alley: https://www.craftscicom.org/ai_signs.html (opens in a new tab), https://www.assertion-evidence.com/ (opens in a new tab).
• SIL, OFL FAQ: https://openfontlicense.org/ofl-faq/ (opens in a new tab). Atkinson Hyperlegible Next: https://github.com/googlefonts/atkinson-hyperlegible-next (opens in a new tab).
• U.S. EPA, Inventory of U.S. Greenhouse Gas Emissions and Sinks 1990-2022, Table 5-11: https://www.epa.gov/ghgemissions/inventory-us-greenhouse-gas-emissions-and-sinks-1990-2022 (opens in a new tab).
• Plan mode and skills in the three tools: Codex (https://learn.chatgpt.com/guides/best-practices (opens in a new tab)), Claude Code (https://code.claude.com/docs/en/permission-modes (opens in a new tab)), Antigravity (https://antigravity.google/docs/plan (opens in a new tab)).