AI for Research Efficiency, In Practice: Make a slide deck that looks like yours
Narration: an AI-generated voice (ElevenLabs).

[0:00] Two decks, one difference
Both decks come from the same paper, agent and model; the one on the right had the presenter's style brief.
The same check, run on both: two hundred and three failures on the left, forty-one on the right.
Anthropic names the pull toward the same look, for web pages: distributional convergence, a model reaching for the most common fonts and colors.
AI for Research Efficiency, In Practice. Make a slide deck that looks like yours.
In this film, a paper and its data become an editable PowerPoint deck in your own style, with one chart that builds in steps, checked against your brief.
Recorded runs in Codex, using four lessons from the series.

[0:50] 1. Ask for the process
Step one: ask for the process. The folder holds the paper, its data, the presenter's website style, the look of this series, and a sample of the presenter's own writing.
The request, in plan mode: 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.
Shift Tab cycles to plan mode, or type slash plan, in Codex, Claude Code and Antigravity.
Codex asks eleven questions, in four rounds, before it writes a plan.

[1:28] 2. Options, in parallel
Step two: options. Three helpers look into the ways to do this, all at once.
The first finds what four vendors do about decks that look alike: each lets you give its tool your own slides, a theme or a brand kit to follow, much like a style brief.
The second compares ways to make a file you can still edit: PowerPoint by hand, a JavaScript library, or Python's P P T X library, which makes real PowerPoint charts with their numbers behind them, and doesn't need PowerPoint.
The third finds how a chart can move and stay editable. Shapes can morph from one slide to the next; PowerPoint's own charts don't, they cross-fade.

[2:14] 3. Attack the options
Step three: a reviewer agent tests each option against the real files, and reports only the problems that would change the result.
Before the run, we listed nine traps we knew were in this folder, and the agents never saw the list. The reviewer found three of them, and part of a fourth.
It also found a problem we hadn't listed: the paper's own figure stacks the closed articles too, up to a hundred percent.
The other five came up at other steps. In its first look, the agent noticed that the presenter's website and video series use different fonts; the website's Lato and Open Sans are both on Anthropic's twenty twenty-five list of fonts to avoid.
And writing the brief, it found the brand's bright orange reaches two point seven to one on its background, where a chart mark needs three to one.
Some university templates need the same check. In one we tested, the first chart color reaches two point six seven to one on white.

[3:19] 4. Decide, and write it down
Step four: decide, and write it down. The presenter chooses the series' font for everything and the bright orange as a tint only, and the agent writes the process as a skill, a file it reads every time.
The brief inside it puts the sources in order, so a disagreement has an answer, and gives each color only the jobs it can do.
For the voice, it keeps a sample of the presenter's own writing, from before any AI tool touched it.
And a dated list of looks to avoid, each with its source, because the looks change.

[3:57] 5. The run
Step five: the run. Check each step at first; once the steps look right, let it run.
It counts the data itself. The file ends its lines the way old Macs did, and its text isn't U T F eight. It reads it as Mac Roman, the old Mac text, and keeps the paper's years, nineteen fifty to twenty fifteen.
Its count for twenty fifteen: forty-four point seven two percent. The paper prints forty-four point seven percent.
The chart that builds is drawn as shapes, each named with two exclamation marks, so Morph can match it on the next slide, where a mask slides back to show the later years.
Then the same numbers as a PowerPoint chart you can edit, with its data inside the file.
Every slide becomes a picture, and a second agent that hasn't seen the code compares them with the brief.
Its report: six problems, from a chart that jumps between two slides to a title that reads like a date. The first agent fixes all six and renders again.
About fifty-six minutes, start to finish.

[5:16] 6. Check it against the brief
Step six: check it against something outside the agent. A script reads the deck file and tests it against the brief.
Text needs four point five to one against its background; large text and chart marks need three to one. Nothing under eighteen points. Every slide gets its own title, and every chart its alternative text.
Then the voice: sentence length, and how often it says I or we, against the presenter's own sample.
One study found that AI summaries of engineering reports made almost no use of the first person, though the reports used it with care.
On this deck: thirty-seven failures, most of them source lines under eighteen points, and titles a screen reader can't find. What it can't see is the data: the chart keeps a test publisher's rows that the paper's code leaves out, which moves the year two thousand by seven tenths of a point.
Two looks are left for you: PowerPoint's own Accessibility Checker, and a last pass in PowerPoint itself, since the preview swaps any font it doesn't have.

[6:31] Same process, another field
The same skill with another person's brief: an extension agronomist, made up for this film, and the E P A's table of methane from rice fields.
Its agent noticed the skill held another presenter's brief, and used this author's own: Lato, field greens, plain words for growers. The deck looks nothing like the first.

[6:57] The line not to cross
Use only fonts, pictures and logos you're allowed to use. An open font can travel inside the deck; a university's logo is usually for its own units.
And if your venue or your institution asks you to say where AI helped, say so.

[7:16] Try it
Pause and try it with the practice folder from the companion page: ask for the best way first. Your agent should ask you questions before it makes a single file.

[7:31] Review and next
So: a brief first, a chart that builds, fresh eyes on every slide, and a check against the brief.
Next: let adversarial reviewers check your paper.