WEBVTT

00:00:00.900 --> 00:00:07.631
Best practices. Instruction files, version
control, careful habits: a whole series of them.

00:00:07.982 --> 00:00:12.790
Very responsible. Too long;
didn't read. Let's just vibe code.

00:00:14.386 --> 00:00:16.668
One request, to a team of agents.

00:00:17.239 --> 00:00:18.750
Twelve and a half minutes later:

00:00:19.261 --> 00:00:23.905
eighty-six years of a river's records,
analyzed, with figures for a paper.

00:00:26.800 --> 00:00:32.290
AI for Research Efficiency.
Episode zero: T L D R.

00:00:33.900 --> 00:00:35.283
First, the quick part.

00:00:35.815 --> 00:00:40.846
Make a folder for your project, right-click
inside it, and choose Open in Terminal.

00:00:41.724 --> 00:00:46.748
Paste one install line:
for Claude Code, Codex or Antigravity.

00:00:47.138 --> 00:00:51.762
Each has its own episode.
Then open the terminal again, the same way.

00:00:52.693 --> 00:00:57.094
Type the agent's name, sign in,
and tell it you trust the folder.

00:01:00.300 --> 00:01:04.113
Now the real part. Say who you are,
and what you want at the end.

00:01:04.444 --> 00:01:06.831
And ask it to spin up a team of agents.

00:01:12.700 --> 00:01:16.946
It works in auto mode: a second model
checks each step before it runs,

00:01:17.097 --> 00:01:19.179
so this run never stopped to ask.

00:01:20.339 --> 00:01:24.326
First, the lead agent set the project up:
a Python environment,

00:01:24.537 --> 00:01:26.190
eighty-six years of gauge records,

00:01:26.360 --> 00:01:31.068
an instruction file it reads every time
it starts here, and a first saved version.

00:01:31.989 --> 00:01:36.718
So those best practices we skipped?
It did them anyway. Show-off.

00:01:38.700 --> 00:01:40.143
Then it split the work.

00:01:40.515 --> 00:01:44.233
Three agents at once:
one on floods, one on low flows,

00:01:44.404 --> 00:01:47.371
and one on the river's whole pattern,
and the rain.

00:01:49.931 --> 00:01:52.942
As each one reported back,
it pulled the results together,

00:01:53.133 --> 00:01:55.882
checked the numbers it would quote,
and drew the figures.

00:02:00.500 --> 00:02:01.241
The finding:

00:02:01.572 --> 00:02:05.469
the river runs about sixteen percent
higher than before nineteen eighty-three.

00:02:05.900 --> 00:02:09.457
By the team's analysis,
more rain explains nearly all of it,

00:02:09.628 --> 00:02:12.773
and its floods and low flows
haven't changed detectably.

00:02:14.273 --> 00:02:17.411
Every stage is saved,
so you can go back to any of them.

00:02:19.500 --> 00:02:21.818
The same approach works on almost anything.

00:02:22.679 --> 00:02:27.224
In this series, a team of agents reviewed
fifty-seven files for a literature review.

00:02:27.655 --> 00:02:32.941
Others mapped a research field, turned a paper
into a narrated video, and built a web page.

00:02:34.713 --> 00:02:36.255
Some call it vibe coding.

00:02:36.786 --> 00:02:41.472
Say who you are, what you want at the end,
and: spin up a team of agents.

00:02:43.700 --> 00:02:48.324
Your run will look different: models update,
and every computer is set up its own way.

00:02:48.754 --> 00:02:51.876
If something goes wrong,
ask your agent to help fix it.

00:02:52.968 --> 00:02:56.415
Before you point it at sensitive data,
watch episode nine:

00:02:56.686 --> 00:03:00.033
your data, your keys,
and your campus's rules.

00:03:00.974 --> 00:03:05.569
And if you ever do want those best practices,
the rest of the series is right there.

00:03:05.941 --> 00:03:06.882
No pressure.

00:03:09.900 --> 00:03:13.368
Next: episode one, why the command line.