WEBVTT

00:00:00.500 --> 00:00:03.949
Both decks come from the same paper,
agent and model;

00:00:04.321 --> 00:00:07.007
the one on the right had
the presenter's style brief.

00:00:07.668 --> 00:00:09.413
The same check, run on both:

00:00:09.846 --> 00:00:13.617
two hundred and three failures on the left,
forty-one on the right.

00:00:14.338 --> 00:00:18.043
Anthropic names the pull toward
the same look, for web pages:

00:00:18.373 --> 00:00:23.419
distributional convergence, a model reaching
for the most common fonts and colors.

00:00:26.800 --> 00:00:32.580
AI for Research Efficiency, In Practice.
Make a slide deck that looks like yours.

00:00:33.900 --> 00:00:39.046
In this film, a paper and its data become an
editable PowerPoint deck in your own style,

00:00:39.276 --> 00:00:42.800
with one chart that builds in steps,
checked against your brief.

00:00:43.633 --> 00:00:47.581
Recorded runs in Codex,
using four lessons from the series.

00:00:50.900 --> 00:00:53.141
Step one: ask for the process.

00:00:53.652 --> 00:00:57.784
The folder holds the paper, its data,
the presenter's website style,

00:00:57.954 --> 00:01:01.776
the look of this series,
and a sample of the presenter's own writing.

00:01:02.736 --> 00:01:06.626
The request, in plan mode:
what is the best way to make a deck

00:01:06.676 --> 00:01:09.132
from this paper that looks
and sounds like me?

00:01:09.644 --> 00:01:12.510
Interview me first,
and don't build anything yet.

00:01:14.611 --> 00:01:21.224
Shift Tab cycles to plan mode, or type slash
plan, in Codex, Claude Code and Antigravity.

00:01:22.225 --> 00:01:26.914
Codex asks eleven questions,
in four rounds, before it writes a plan.

00:01:29.300 --> 00:01:35.058
Step two: options. Three helpers look
into the ways to do this, all at once.

00:01:36.119 --> 00:01:39.807
The first finds what four vendors
do about decks that look alike:

00:01:40.238 --> 00:01:42.614
each lets you give its tool your own slides,

00:01:42.804 --> 00:01:46.311
a theme or a brand kit to follow,
much like a style brief.

00:01:47.452 --> 00:01:52.195
The second compares ways to make a file
you can still edit: PowerPoint by hand,

00:01:52.365 --> 00:01:56.117
a JavaScript library,
or Python's P P T X library,

00:01:56.287 --> 00:01:59.238
which makes real PowerPoint charts
with their numbers behind them,

00:01:59.409 --> 00:02:01.189
and doesn't need PowerPoint.

00:02:02.290 --> 00:02:05.632
The third finds how a chart
can move and stay editable.

00:02:06.102 --> 00:02:11.945
Shapes can morph from one slide to the next;
PowerPoint's own charts don't, they cross-fade.

00:02:14.900 --> 00:02:19.413
Step three: a reviewer agent tests
each option against the real files,

00:02:19.564 --> 00:02:22.551
and reports only the problems
that would change the result.

00:02:23.772 --> 00:02:27.457
Before the run, we listed nine
traps we knew were in this folder,

00:02:27.607 --> 00:02:32.994
and the agents never saw the list. The reviewer
found three of them, and part of a fourth.

00:02:34.295 --> 00:02:36.437
It also found a problem we hadn't listed:

00:02:36.747 --> 00:02:41.652
the paper's own figure stacks the closed
articles too, up to a hundred percent.

00:02:43.212 --> 00:02:45.334
The other five came up at other steps.

00:02:45.824 --> 00:02:46.815
In its first look,

00:02:46.945 --> 00:02:51.518
the agent noticed that the presenter's website
and video series use different fonts;

00:02:51.909 --> 00:02:53.100
the website's Lato

00:02:53.150 --> 00:02:58.194
and Open Sans are both on Anthropic's
twenty twenty-five list of fonts to avoid.

00:02:59.151 --> 00:03:03.378
And writing the brief, it found the brand's
bright orange reaches two point seven

00:03:03.428 --> 00:03:07.415
to one on its background,
where a chart mark needs three to one.

00:03:08.898 --> 00:03:11.505
Some university templates
need the same check.

00:03:12.057 --> 00:03:17.472
In one we tested, the first chart color
reaches two point six seven to one on white.

00:03:19.700 --> 00:03:22.824
Step four: decide, and write it down.

00:03:23.255 --> 00:03:28.251
The presenter chooses the series' font for
everything and the bright orange as a tint only,

00:03:28.722 --> 00:03:33.487
and the agent writes the process as a skill,
a file it reads every time.

00:03:34.788 --> 00:03:39.252
The brief inside it puts the sources in
order, so a disagreement has an answer,

00:03:39.642 --> 00:03:42.424
and gives each color
only the jobs it can do.

00:03:43.424 --> 00:03:46.812
For the voice, it keeps a sample
of the presenter's own writing,

00:03:46.942 --> 00:03:49.467
from before any AI tool touched it.

00:03:50.548 --> 00:03:55.950
And a dated list of looks to avoid, each
with its source, because the looks change.

00:03:58.100 --> 00:04:04.708
Step five: the run. Check each step at first;
once the steps look right, let it run.

00:04:06.309 --> 00:04:07.950
It counts the data itself.

00:04:08.340 --> 00:04:13.253
The file ends its lines the way old Macs
did, and its text isn't U T F eight.

00:04:13.624 --> 00:04:16.715
It reads it as Mac Roman, the old Mac text,

00:04:16.926 --> 00:04:20.928
and keeps the paper's years,
nineteen fifty to twenty fifteen.

00:04:22.076 --> 00:04:26.102
Its count for twenty fifteen:
forty-four point seven two percent.

00:04:26.493 --> 00:04:29.457
The paper prints forty-four
point seven percent.

00:04:31.212 --> 00:04:35.935
The chart that builds is drawn as shapes,
each named with two exclamation marks,

00:04:36.185 --> 00:04:41.467
so Morph can match it on the next slide, where
a mask slides back to show the later years.

00:04:44.608 --> 00:04:49.673
Then the same numbers as a PowerPoint chart
you can edit, with its data inside the file.

00:04:51.165 --> 00:04:52.972
Every slide becomes a picture,

00:04:53.183 --> 00:04:57.238
and a second agent that hasn't seen
the code compares them with the brief.

00:04:59.459 --> 00:05:04.909
Its report: six problems, from a chart
that jumps between two slides to a title

00:05:04.959 --> 00:05:09.807
that reads like a date. The first
agent fixes all six and renders again.

00:05:11.350 --> 00:05:13.932
About fifty-six minutes, start to finish.

00:05:17.300 --> 00:05:20.871
Step six: check it against
something outside the agent.

00:05:21.283 --> 00:05:24.673
A script reads the deck file
and tests it against the brief.

00:05:25.914 --> 00:05:29.137
Text needs four point five
to one against its background;

00:05:29.567 --> 00:05:34.081
large text and chart marks need three
to one. Nothing under eighteen points.

00:05:34.531 --> 00:05:38.854
Every slide gets its own title,
and every chart its alternative text.

00:05:41.394 --> 00:05:43.579
Then the voice: sentence length,

00:05:43.689 --> 00:05:48.379
and how often it says I or we,
against the presenter's own sample.

00:05:49.289 --> 00:05:56.134
One study found that AI summaries of engineering
reports made almost no use of the first person,

00:05:56.184 --> 00:05:57.144
though the reports used it with care.

00:05:58.564 --> 00:06:04.028
On this deck: thirty-seven failures, most
of them source lines under eighteen points,

00:06:04.259 --> 00:06:06.490
and titles a screen reader can't find.

00:06:06.861 --> 00:06:08.612
What it can't see is the data:

00:06:08.982 --> 00:06:12.855
the chart keeps a test publisher's rows
that the paper's code leaves out,

00:06:13.146 --> 00:06:16.608
which moves the year two thousand
by seven tenths of a point.

00:06:18.386 --> 00:06:22.129
Two looks are left for you:
PowerPoint's own Accessibility Checker,

00:06:22.400 --> 00:06:27.865
and a last pass in PowerPoint itself, since
the preview swaps any font it doesn't have.

00:06:31.700 --> 00:06:36.884
The same skill with another person's brief:
an extension agronomist, made up for this film,

00:06:37.134 --> 00:06:40.196
and the E P A's table of
methane from rice fields.

00:06:42.356 --> 00:06:47.319
Its agent noticed the skill held another
presenter's brief, and used this author's own:

00:06:47.769 --> 00:06:53.573
Lato, field greens, plain words for growers.
The deck looks nothing like the first.

00:06:58.100 --> 00:07:01.606
Use only fonts, pictures
and logos you're allowed to use.

00:07:01.936 --> 00:07:07.805
An open font can travel inside the deck; a
university's logo is usually for its own units.

00:07:08.849 --> 00:07:13.708
And if your venue or your institution
asks you to say where AI helped, say so.

00:07:17.300 --> 00:07:22.774
Pause and try it with the practice folder from
the companion page: ask for the best way first.

00:07:23.085 --> 00:07:26.814
Your agent should ask you questions
before it makes a single file.

00:07:31.600 --> 00:07:34.606
So: a brief first, a chart that builds,

00:07:34.957 --> 00:07:38.523
fresh eyes on every slide,
and a check against the brief.

00:07:41.200 --> 00:07:44.810
Next: let adversarial
reviewers check your paper.