That moment in time, I'd never launched

anything on GitHub. I'd never done a

solo project at all, and I launched it,

and I made a video, and it went viral,

got a million views. That's how I got

started.

>> Did you know [music] if it was actually

possible to bring in all of these

signals before actually building the

thing?

>> Not not a clue at all. I fundamentally

believe that very [music] soon no humans

should ever write any code, and that no

humans should ever read any code. Had

the audacity to start a programming

language at 5:00 in the [music] morning

that Sunday morning, and it was a stupid

idea, but it led to a good idea.

>> Matt Van Horn, thank you for joining

Human in the Loop.

>> I I can't wait to be removed from the

loop after.

>> Yeah, [laughter] exactly.

>> to you're going to have to rename your

podcast soon, Alex, cuz I I I don't want

I don't want to work work on this

anymore. I just want my agents to do it.

>> Honestly, Human Out of the Loop is like

an easy rebrand, so we'll do that at

some point. Um it's awesome to have you

you know, you're you've become like a

prolific builder and engineer, and I

think what's most fascinating about you

is you're not technically trained,

you're not an engineer by training, yet

yeah, I feel like you're doing circles

around a lot of classically trained

engineers. And so, my goal is just to

understand

how do you build, and and how were you

able to work up the curve of

understanding all of this

so quickly? Um and so, yeah, I just want

to watch you cook. Obviously, you've

learned a lot in this process, but also

there's it seems like there's a ton of

stuff you don't even know what it means

or or kind of the exact details as

you're building it, but that's also kind

of the beauty of this technology. It's

like if you just have enough agency and

you're you're dangerous enough to ask

the right question, that's kind of all

you need.

>> Yeah, it's it's all about confidence in

asking the right questions.

>> Yeah.

>> Right? Like like the fact that I had the

audacity to start a programming language

at 5:00 in the morning that Sunday

morning was stupid.

And it was a stupid idea, but it led to

a good idea.

>> Totally.

>> And and so I think just not being afraid

and just starting is is really magical.

And so like Pretty Press, like you're

talking about programming languages and

so I've gotten contributions into Go,

which is

incredible programming language. Um

and but Peter makes all of his CLIs in

Go. And so it's like so

uh Peter Steinberg, so I was like,

"Okay, I guess the Pretty Press CLI

should be in Go." Right? I've never done

a project in Go before. And what's funny

is I I built this other project Agent

Cookie, which we don't have to spend

much time on, but it it kind of

automatically syncs your Chrome cookies

from your main Mac to your Mac Mini in

the cloud.

>> Mhm.

>> And I literally was about to ship this

product and I have I don't I don't ever

look at code. Um

and [snorts] I I philosophically don't

believe in looking at the code. And

>> Well, why is that?

>> That's that's that's for agents. That's

agent's job. That's why I don't have an

IDE. Like why do I don't need to look at

the code. Um I can talk to my agent

about a problem, but I was literally

shipping this product and I was on the

GitHub page and I was like, "Oh, I had

no idea Agent Cookie was built in Go."

Like I literally didn't know that it's

94.7%

>> [laughter]

>> Go.

>> Do you think Do you think your view on

not needing to read your code is

specific to like

these side projects that you're building

or like do you believe that even in the

context of building a company?

>> I fundamentally believe that that very

soon no humans

should ever write any code

and that no humans should ever read any

code.

I strongly believe that.

And um and again, I get to personify

this because I can't read the code. So

like I I already have this this problem

that I literally would not be useful,

right? But I I think the the the

engineers of the future, like what is an

engineer? An engineer is a problem

solver.

It's not someone that just you know,

studied the right things. Right? An

engineer a solves problems. Like we we

tell our team that your your job is to

automate your job and find something

else more important to work on, a more

important problem to solve.

And so that's that's something I I

strongly believe in.

>> Super interesting. We may have to do a

episode in the future where we do a

debate style episode cuz I feel like it

would just be fun. There's for sure

people who would take the other side of

the trade and I think it would it just

like the intellectual exercise of

understanding why or

why or why not would understanding the

code and reading the code in the future

be valuable. And my assumption is your

argument is the models are going to get

so good that even if people are working

about worrying about scalability or like

how hard in the code is, these are all

solved problem by good enough models.

>> Yep.

>> Yeah, makes sense.

Um okay, let's let's build something.

What what do you want to build?

>> Let's let's build a few things. I have a

few few things running. Let's start

Let's start here. So so Trevin who I

built Pretty Press with, I asked him,

"Hey, what what should I build on on air

today?" He's like he's like

nutritionvalue.org. And I'm like, "What

is that?"

And I'm like, "This This is this looks

horrible. Um

and um

And he's like, "Yeah, it's not pretty,

but it's really valuable and can pull in

lots of valuable information."

And so

the I don't know if it's going to work,

but we're going to try to build a CLI

about

nutrition information.

>> So basically this website is just like

this mega database of the nutritional

facts of a ton of different foods and

food items.

>> Yep.

>> Cool.

>> Yep, you're you're seeing my

nutritionvalue.org password as I

generate it. Uh-oh. [laughter and gasps]

Try not to try not to steal my secrets

of my nutrition value.

>> All right, I'm in. Okay, so

what what I'm going to do is I'm going

to make a new window here.

And so I I use this skill um

Uh sorry, I'm going to use I'm going to

use pretty press. All right, so CLI

printing press.

Generate a a ship uh ready CLI for an

API with lean research generate build

ship check loop.

So, I don't know why I have two.

Hopefully, it will.

>> And so, I I know uh while while this is

going, I know like I read your agent

hacks article and you talk about like

the importance of you use the compound

the C plan um like you you use compound

engineering for a lot of what you build.

Is that just implicitly built within the

printing press, so it goes through that

process?

>> It's it's actually not, but it's

obviously heavily inspired and similar

to and learned from all the best

practice. So, it doesn't actually run

compound engineering, but

Trevin is one of the best agentic

builders that I I've met and uh a lot of

the best practices that him and Kieran

have built into

compound engineering are built in

obviously to the printing press.

And so, I'll start the printing press

for nutrition by first running the

mandatory pre-flight setup. So, I don't

know what it's doing, but it'll it'll

tell us I mean, why don't we go to

another window and try try something

else? I think I can show you something

else I was working on. I can show a

compound engineering thing. Okay, so

we we have a um

a

a CLI in the printing press called the

ESPN one.

And so, I said

uh

earlier, "Why did England win yesterday

in the World Cup?" And I'll look up

England's World Cup match from

yesterday. First, verify the CLI. So,

it's using the ESPN CLI that that we

built. And so, came with the result.

England beat DR Congo 2 to 1 in the

World Cup round of 32. Yeah, the short

answer is they won is Harry Kane with a

big assist literally twice for

uh from substitute Anthony Gordon. It

was actually a near disaster comeback

rather than a dominant win.

And um

so, what's what's interesting is this is

the second So, when we built the ESPN

CLI, there was no mapping to

the World Cup cuz it hadn't been built

yet in ESPN system. And so, I feel like

this took longer than it should have. It

took 1 minute 11 seconds. And so, one of

the things that the features that were

working on right now in the printing

press is self-learning CLIs. So, the

agent will leave notes for itself for

next time locally.

And ESPN's one of the first ones to have

this feature. But,

I just ran this in another context

window, and it also took 1 minute 10

seconds. So, this one should have been

faster. And so, I've no idea what the

result is going to be, but I'm just

going to say I wrote this right before

we started. See, E plan, this took 1

minute 10 seconds and 11 seconds, and I

pasted in I'll do a lot of copying and

pasting, a lot of screenshot pasting. Um

Hey,

this was after the search deal. So, this

is the the newer context window. Why did

it just as long as what just happened?

So,

All right. So, we got got that one. And

>> And basically and basically your whole

thought is this should have been faster

because if you have this self-improving

CLI, it should already have had the

information about the England win.

>> Exactly. It should Not the England win.

It should have known how to map a World

Cup

>> Got it. Got it.

>> the other question I asked was about

Belgium

and so, it should have learned the

mapping. Kind of again, CLIs, think of

it like Google Maps for things. So,

while that's going on, I think our CLI

had a question.

Um

So, it's it's asking, "Anything you want

me to know before I begin? A vision for

the CLI should do some feature care

about or off." And it's like, "Let's

just go. I'll start research now. I'll

ask about API keys, browser auth, other

things. Cool."

So, I don't have a particular vision for

this. I I hope it's self-explanatory.

>> I also know that you feel pretty

strongly about um

uh dangerously skip or bypass

permissions. What Why do you feel

strongly about that versus needing to to

check and keep yourself in the loop?

>> So,

there's just a lot of stupid stuff that

comes up if you don't have it on. So,

like for example,

it's going to go So, the what what the

printing press does right now is it's

going to go research

nutritionalvalue.org. It's going to look

around. It's going to do a hard sniff to

try and figure out kind of the mapping

of the whole site. It's also going to go

on GitHub and look for community um

CLIs. Like, has anyone done this before?

Like, when we were printing the Domino's

Pizza CLI, someone had a project that

was a a Python script. It's like pre-AI.

A Python script to be able to order

Domino's Pizza. Like, they sniffed the

APIs from domino's.com and they just

wanted thought it they wanted to run it

in Python. Project didn't have that many

stars. I think it was less than a

thousand, but like it worked and was

really cool.

Uh-oh, nutritional value or explicitly

prohibits programmatic access to data

and has no official It largely comes

from USDA food data, which offers a

official free API with nutrition. How do

you want to proceed?

>> So, what are you going to do?

>> I don't

Let's try building a combo CLI. So, try

and do that one, but also get the the

one that has an API. But, I I do like

nutritionalvalue.org

but I like that one. So, let's try and

do both. YOLO mode.

>> Is it basically what it's saying is that

nutritionalvalue.org

like generally prevents you from doing

any sort of automated behavior? So, so

you have to decide whether you want to

still try to

have have Claude code or the the

printing press figure out how to still

deal with that?

>> Yep, exactly. So, combo CLI it is. I'll

treat nutritionalvalue.org as a primary

source headline commands top of the read

me first run experience USDA as a

secondary.

Let's go with Piers no primary. All

right, so

Now, am I in trouble with Fable 5

safeguards flagged and safeguards are

intentionally broad right now.

These measures let you bring I don't

know what's going on. Hopefully it still

works.

>> Did they change the model or is it still

using Fable?

>> I'm not sure. So anyway, so that's

that's going in there. So now let's go

over here.

All right. Short answer, the learning

that got stored was the wrong kind to

help the second question. The kind that

would have helped was never recorded and

even a perfect hit would have maybe

shaved maybe a third of the time.

What got caught was a per game fact and

the second question was about a

different game. The Belgian session

taught this query maps to this event. My

England session run recon on the CLI

correctly rejected the learning. The

transferable knowledge was never saved.

What both sessions actually rederived

from scratch was the method. The World

Cup league code is soccer.fifa.world.

The date flag is {dash} {dash} dates not

{dash} date.

Even a warm hit would have made it feel

fast.

>> So basically it's just saying that it

didn't properly map how to get to World

Cup's

information quickly.

>> Yep. So the fix now applied. So I

recorded a playbook for the query

family. And so I'm actually going to

yell at it right now. So this will be

fun. So I'm going to start with a CE

plan.

I don't want you to fix this for me.

What I actually want you to do is I want

you to submit a PR to the printing press

library for the ESPN CLI so that it can

learn not how to find the World Cup but

how it can learn to self-heal itself in

the future when something like this

happens.

And so the that's the promise of a

self-healing CLI and clearly you failed

and what you've built so far is not

working out. So I need you to go deep

and figure out how to systematically

programmatically

fix this not just for ESPN and future

events. Could be a tennis event. Could

be something else. This isn't about the

World Cup. I think this is a good

example because when we did the original

HAR sniff, the World Cup wasn't

happening and so it doesn't know how to

map to the World Cup. But the whole

point of a self-learning CLI is to do

that. So there's two paths I want you to

figure out here. One is what should be

the change to the ESPN CLI, but two this

we want to build a self-healing CLI into

every single press, everything that

happens within the printing press. So I

need you to look at the actual printing

press code to figure out how we can do a

better job here.

>> This is like Matt Van Horn ASMR leaning

into the mic

talking to to your agent. That's great.

Um one question, what what is

what you use the the phrase sniff HAR

sniff what what is that referring to?

>> Yep. Uh by the way, it's accidentally

building a PR for that thing we said

before. I don't know.

I don't think I asked it to start

writing code and building it, but this

is that open CV thing. There's a plan

written, there's a fix committed.

There's a build verified. Okay.

>> So it's actually submitting a PR for

whatever that bug was. That's

>> Apparently I I literally was not using

my normal open source software stack. So

it was just doing things. I'm going to

I'm just going to skip that one right

now.

>> Okay.

>> Um

all right, the US needs a free API key.

API.data.gov no key is set up in your

environment. How should I handle this?

Um

>> So what which which one is this? Which

job is this that we're working on right

now?

>> Sorry, am I This is now we're we're on

the the printing press one. So I need to

sign up for an API key for API.data.gov.

How do I get one of these?

It's a free for federal agencies. I'm

not a federal agency. I'm so confused.

I'm not a federal agency. How do we get

one of these?

Um so

>> way, that I don't know if this is like

possible, but the this it feels like

such a thing now when you whenever

you're building stuff now it feels like

you constantly have to go and retrieve

API keys. Like we as the human are just

like the

the the errand boy of our agents to get

API keys. I've been thinking now like,

is there ultimately going to be a way

easier way to get API keys than us

having to go to sites to get them for

our agents?

>> Yes, soon. They're better be.

>> Yeah.

>> So, here's here's another window I want

to kick off going, well, so

this is my start I sent this message

right before we started talking. So,

this is a new window to this context of

this conversation. I said, I made a PR

to for self-improving CLIs that agents

leave notes for future agents, did it

merge? What's the status? Cuz I know

Trevor was supposed to look at it.

And I didn't remember if it had shipped

or not.

And he's like, it merged. So, great. And

I was like, why haven't I seen it? I was

like, okay.

Is it working? Can you see all the So,

theoretically, all CLIs, new ones

printed after this point, should um

should um

have it.

And it said, no.

>> Mhm. And so, in theory, if this had

worked, if we go back to the uh ESPN and

World Cup example, you should have been

able to very quickly just get

information on the Belgium or England

game because the ESPN CLI should have

through self-healing mapped how to get

the World Cup information.

>> Exactly.

>> Yeah.

>> And so, I had hand-built ESPN in the

prediction go with self-learning.

But, I wanted everyone that makes a new

one to also have the self-learning. And

so, I'd spent a lot of time with the

ESPN one working on that. And so,

anyway, but here's the thing.

Apparently, what I built never pulled it

all together.

>> Oh, no.

>> Adoption is zero. I code searched the

whole thing.

The reason it's gating, the loop is

opt-in for the spec.

Apps are disabled with a benign go

no-go. And here's the gap, it just never

mentions enabling it.

So, now I'm going to say,

I'm doing a live demo right now, and

this is making me look really bad. And I

said that I was working on this

this self-learning CLI feature, and that

you told me it was merged and live, but

apparently it's not. I want this

default on, and I want it to be high

quality. So, if there's anything that's

not high quality about this, any reasons

it should be off, like I want every new

CLI that's printed to be able to have

this. And if the product's not good

enough, then we should talk about that.

So, I want you to do a deep dive into

this and figure out A, how we turn this

on, and how we start printing things

with self-learning turned on, so that we

can learn from actual data of new CLIs

that are printed in the printing press

if this is a valuable feature or not. I

think it's extremely valuable, and if

it's not valuable now, then we need to

make it valuable. So, I believe in you.

Go, Agent, go.

>> [laughter]

>> By the way, what do you use for

for yap to text? I love the little icon

in the bottom right that shows as you're

talking.

>> Yes, so this is

um

uh Monolog.

>> Cool. Oh, love it. Love it.

>> And so, all right, let's see. We've got

other stuff going on here.

I have the grounding any, the press

already does this. Okay.

So, let's make it a plan there.

>> So, basically, just to reorient, the

things we're working on right now is we

have the nutrition, whatever it is,

nutritionfacts.org,

which we were trying to create the print

it the new printing press CLI for, so we

can pull nutritional information. You

basically hit a wall there cuz it was

like, "Ah, you can't programmatically

pull information from the site." But

then you were like, "Let's just build a

combo CLI that's primary from nutrition

facts, and then from USDA if you can't

pull from that." That's one. The second

is the ESPN CLI and making it

self-healing, so it pulls the right

information. The third it was this

OpenCV bug that you somehow were fixing

even though you didn't know.

>> Yep. I I just closed that for fun cuz I

it wasn't running my right stack.

>> Cool. And then, is there anything else

we had going?

>> I think that might be it. So, it's it's

struggling with that that nutrition

value thing, which is uh this is a good

demo cuz it shows that some stuff does

not

uh want to be safe. So, anyway, I'm

going to browser sniff it anyway. So, I

could say that it could hand author an

HTML spec.

So, endpoints from the structure I

already mapped. Fast, no browser data,

standard HTTP, browser sniff it anyway.

So, let's try.

>> What does browser sniff mean? Is it just

literally studying the site?

>> Yeah, so it it's

like

um way to think about it is

uh when you load a webpage, you click a

link, or something happens, um it

there's a it's calling APIs on its own.

And sometimes

um

they people want to

hide them. Like LinkedIn loves hiding

their stuff because people like to

scrape their content and things like

that. Kind of bot abuse type stuff. But

like a site like this, like I don't know

I don't know why it's why they care so

much. Um

And so, like and here's one that makes

me really mad. I I love uh

I love Hyatt.

>> Mhm.

>> I'm like

uh

always trying to get globalist every

year. And uh

they have so much security on Hyatt,

it's extremely frustrating cuz all I

want to do is spend more money and spend

more points at Hyatt and have my agent

be able to book me hotel. And they have

some of the highest security

>> it's like Fort Knox. Like you cannot

build a CLI to Hyatt.

>> It exactly. So, anyway, it's it's trying

now to to do this over here.

All right. Confirm the scope and I'll

proceed. Um

force it on all prints.

Yes.

>> So, this was the the self-healing, just

making sure it's applied to everything

printed from the printing press.

>> Yep. Exactly.

>> Cool.

>> And so, that that's running there.

Um this one's analyzing the learn loop

pattern. So, I think this is our

football one. Yeah, this is our football

one. So, it's kind of trying to figure

out what's going on with with the ESPN

here.

>> What one question just as these are

running is um and and by the way as

we're talking through this if there are

any other agent agentic hacks that we've

missed that you think are important to

call out feel free to call them out on

the compound engineering piece like why

do you why have you liked compound

engineering the process so much and like

can you just roughly outline like the

way that it works like if you were to

build something from scratch that isn't

one of your existing tools right now

what would be kind of the sequence of

using compound engineering?

>> Yeah, so the moment you have an idea

make a a CE plan.md. And so

the the basic gist is is agents are

inherently lazy like they want to work

as little as possible to make you as

happy as possible. It's kind of like

that that that tricky balance. And

something that I found that's extremely

valuable is to to force the agent to

write itself a plan file. So think of it

like like a source of truth a a text

file that it's always going to go back

to to make sure that it doesn't like

lose track of things cuz agents can

either end things abruptly when they get

lazy or they end up on a tangent and

they kind of forget their original

purpose.

And so compound engineering I think does

the most thoughtful job of writing a

high-quality plan.md file

for your agent. And I think one of the

biggest shifts cuz

I've sorry I've written two of these

agentic engineering hacks articles and

the the first one says like make a plan

and then use this editor to read the

plan and modify it. And the newest

version says never ever read the plan.

Right? So don't read the plan.md.

[snorts]

I always make the plan.md. I almost

never read it. Plans are for agents you

silly human.

>> So it's the same reason you say don't

read the code. Same idea.

>> Exactly. So again this is like how my my

change and philosophy has shifted over

time. So the the first is like should

never read the code. Now it's never read

the plan. And you can uh ask questions

of the

>> Next it's going to be just never read.

In life, never read.

>> Again, [laughter]

like I I have a problem with the name of

your podcast because I don't want to be

in the loop. Like fire me. Please. Like

like I just

I want to be I want to be removed from

this uh

which is uh and so yeah, so kind of just

again, how I started with I want to

build a programming language from

scratch was CE plan and then using my

voice was just like

I have this crazy idea. I don't know if

it's good. Tell me good ideas, bad

ideas, do research. CE plan is also

very, very good at

at research and kind of searching the

internet and like again, back to

dangerous skip permissions. Like

it

uh without dangerous skip permissions

on, it's like, is it okay if I go to

google.com and look for something?

Is it okay if I go to this webpage?

Like, what? Yes. What are you doing?

>> exhausting, yeah.

>> Like like we need to we need to move.

Um

and so it's see what I even know what

This is the self-improving CLI.

Um

Okay, so this is the one where it's

trying to fix the overall

>> Yep.

>> one there. This one is doing something

with nutritional values.

Their internal IDs are for USDA IDs and

it is data is USDA data, but NV So

anyway, it's

it's learning something.

>> And and again, when people hear

self-healing CLI again, I think for like

a non-technical person, they're like,

that sounds super technical.

Self-healing, what what does this mean?

Is the self-healing CLI literally just

like it is a markdown file that has

notes about previous learnings with the

CLI that the CLI reads before it takes

action to kind of know where it should

go?

>> Exactly, yeah. So the the the

Yes, you you've nailed it, but the the

the clearest example of what this world

like when we built the ESPN CLI, it

didn't know the World Cup

match patterns, right? It knows how to

find NBA games. It knows how to find MLB

games very quickly. That pattern has

been that way for a long time.

It has no idea how to find a World Cup

game because when it was taught,

it didn't know. So, an agent is fumbling

around looking for, "Okay, where's the

World Cup? How do I find this?" and and

go from there.

>> Cool.

>> And then let's see what else is fun in

the article. So, so this is a good one.

So, you CE plan for your deepest

non-engineering

um work, make a plan for the plan. So,

this this is really magical. So, the the

example I gave in the article here was I

was meeting with uh Michael Mauboussin,

uh who used to run uh was the research

partner at Google Ventures.

And so, kind of very focused on like

product market fit for for businesses.

And one of the things he said was, "Oh,

you should like read my book. It's uh

it's would be great to to think about

your your challenge and problems." And

so, what I did, so I I Granola

everything. Uh so, Granola's one of my

favorite pieces of software. It takes a

transcript of any conversation that that

you have.

And you can copy that whole thing into

your your agent and use that.

>> Yep.

>> And so, in our whole meeting with

Michael, I was like, "Hey, can I can I

Granola this? I would love to like jam

with you on product market fit stuff for

for my business."

And so, we had a really good

conversation. It was like a 2-hour

conversation. And then he was like, "You

should read my book." And so,

in an agentic world,

like I want to take the context that I

have from the 2-hour meeting. So, I took

the entire 2-hour Granola. So, I did CE

plan, make a plan for the plan

um

of I want you know about my business. I

want to work on product market fit uh

and a plan. Um I want you to in a

non-lazy way read Michael's book. And I

I you to take the context from this um

meeting we had from him to create a plan

of

of how how we work together on this

business problem.

And so, what it did was because it wrote

this plan on MD file,

um

and by the way, this process was the

longest research pro- process I've ever

seen Cloud Code do. It spent about an

hour on this.

>> Wow.

>> And what it did is it read every single

chapter and wrote it required itself I

didn't come up with this the agent came

up with this to write itself a book

report about each chapter

and specific to my business.

And

>> Crazy.

>> that and it put together this crazy plan

that was

really wild. And another example of

something kind of a very

a normie thing to do was to use um

to use

uh CE plan to plan a Disney World trip.

So, the first thing I did is I ran last

30 days on Disney World to see like,

"Hey, is there anything new that's

opening soon? New firework shows? New

New things like that?" To kind of just

get the the download of like the latest

stuff from Reddit X, what people are

talking about, what are the longest

lines, what secretly has the shortest

lines, where should you go at rope drop?

Like these sorts of like important

social signals. And then I was like,

"Okay." And I was like, "I have a

10-year-old, an 8-year-old, 5-year-old,

and 2-year-old. I need you to map out

what rides everyone can go on and can't

go on. I don't know their heights. I

don't know how many inches they are. So,

you have to make assumptions. Assume

The two older ones are tall, the others

are average." Um and uh

go. And it asked me a few questions and

then it created this whole thoughtfully

designed

uh

web page uh which was really really

nice. And it also gave me instructions

to give to my Open Claw

of the reminders is like, "By the way,

like 7 days before your date in the

park, you need to sign up for this pass

to like skip the lines and it opens at

like 6:00 a.m. East Coast time.

Um you should add it to your calendar

and copy this to your um

your open clause so that it it will do

this automatically.

>> That's awesome. I love that. And and

just thinking about the alternative,

like if I think about that Disney trip,

if you did not have CE plan and if you

did not have last 30 days,

you know, the average person would just

go type into

uh Claude, co-work, Claude code,

whatever, chat. Um they would basically

say, "I'm going to Disney World with my

family. Build me a plan. The the I have

four kids. The these are their ages."

Like they give the same context, but

my assumption is it would not be nearly

as tuned to what matters now at the park

versus what generally matters. And it

would not be tuned to like also context

about you in how this uh this plan to be

executed on given kind of the agentic um

workflows you already have like at your

disposal.

>> Yep.

Yep.

>> Yeah. So they By the way, our last 30

days result came in for your World Cup.

>> Oh, nice. What came in?

>> So Paraguay knocking out Germany is the

story of the tournament so far. Paraguay

held Germany 1-1 through extra time and

won the shootout 4-3.

Germany's first ever World Cup penalty

shootout loss.

Um

and then, let's see, Jen someone to

break down what went wrong for Germany.

Um let's see, US Ventis finally won a

knockout game and fans are euphoric and

salty about the ref.

Um

and so it's pulling in from Reddit.

World Cup thread on red card rules

mainly for the Balligan situation blew

up with 3,000 upvotes and 5,000

comments.

"They forgot to add the referee to the

Bosnian starting lineup," quipped Goofy

Graffer.

Um England versus Mexico is the next

blockbuster and DR Congo won hearts on

the way out. This is awesome.

Uh the top comment

on this video was pure joy worded here.

We lost, but I'm proud of my country. We

tried. Congo Congo emoji with 5,700

likes. Uh England vs. Mexico about to be

a blockbuster movie.

Money says France with Argentine right

behind it. The Polymarket World Cup

market has France as the clear favorite,

roughly 33% after their free around 32.

Argentina around 22%.

And there's even a live market whether

Declan Rice wins the Golden Ball.

And Messi to win the Golden Boot with

44%.

This is awesome.

Yeah, now what's so interesting is like

I'm just thinking about all the

different ways that last 30 days, like I

know this is just like a a side project

for you, but I'm just thinking about

like there are actually so many ways it

could be commercialized. Like I think

about like um you know, Greg Eisenberg

has like his startup ideas finder type

thing. Like I could imagine that type of

thing, but using last 30 days with all

these signals about the biggest problems

people are experiencing

um in a post-AI world, whether

personally or professionally. And then

also I think about like how big the

market is for social listening for

companies. And how good this would be

for companies doing social listening

into their consumers.

Yeah. Totally.

>> Very cool.

>> Okay, so here we go. It's

Nutrition PPCLI absorb manifest.

So, 15 absorb table stakes features and

seven novels, so 22 features for this

CLI. The field is wide open. No mature

USDA CLI exists. Best competitors under

five stars. Nutrition value has almost

no tooling. This shifts past every eight

competitor with offline caching, agent

native output, and cross-source

enrichment nearly no other tool has.

The architecture, it's an aggregator

combo. It uses USDA's open API seeds.

Then generated baseline search get batch

and SQLite store. Nutrition value is the

hand-authored HTML so that both feed one

unified food store keyed on source and

ID.

Since the keys are the same FD

ID is cross-reference to be exact. The

seven novel features all hand code

granted in three personas. So literally

it came up with three personas of power

users of this nutrition app. So it

pretended to be Marcus the power lifter,

Dana the low-carb dieter, Priya the

agent builder to think about what their

use case would be and why they want this

CLI.

>> So crazy.

>> It came up with enrich. Merges nutrition

values derived analytics and that carves

omega-6-3 database percent onto a USDA

record. The Helen feature numbers no

USDA API call returns resource with only

two scrapers total.

Rank top bottom foods by any 16

ingredients category data set

filterable. Compare 1,000 kcal two to

five foods side by side per 100 g per

serving. No existing nutrition CLI

offers a comparison basis at all. Find a

minimum protein 20 compound

multi-nutrient threshold discovery.

Marcus is high protein under 165 kcal in

one step. This is blowing my mind by the

way.

>> Crazy crazy.

I also just don't understand how I don't

actually understand how this pulled it

off when it started by say two things

that felt like blockers. One is it said

it couldn't do any sort of like

programmatic pulling of data. So I don't

understand how it got around that. And

then second, you had to get an API key

from like a government website which you

didn't do. So how did it get all this

information without doing those things?

>> It hasn't built it yet. It's about to

build it. All right?

But here's what it says. Things to note.

No stubs. Everything list fully

implemented. Pure balance NV leads the

top two novel commands and powers the

original USDA backs the rest. NV can't

have more commands cuz its value is an

overlay and shared ID data not a

separate feed.

Nutrition value terms of service. It's

homepage prohibits scripted access. You

approved building against it. Yolo

[laughter] mode. The CLI stays polite,

honest UA, conservative rate limiting,

no crawling, and the read me will

document this copy. So, it's not going

to be super aggressive. So, it's built

it's it's it's planning to build its own

rate limiting

>> Yep.

>> into it because it knows it's not

allowed even if it has top blockers.

>> Uh seven candidates were cut. This

adversary past killed anything on ver

So,

>> That's so crazy.

>> approve, generate now.

I have ideas to add, trim scope, review

full manifest file. So,

let's approve.

>> Crazy. And like

I know this it we're like demoing this

now, but like this is how you would go

through the process of printing a new

say a say a light, right? This is like

exactly how you do it.

>> Absolutely.

>> Cool.

>> If there there is there is still other

process.

>> Yeah.

>> And if And if I'm unhappy with something

here, I would then literally copy this

whole thing.

And I like I I would do this, but I'd be

like, "Hey,

this is called the Yolo printing press.

I have Yolo mode. There's no Yolo mode.

Like Like don't ask me dumb questions on

if I want to scrape it or not. Like

That's how I would literally I could I

could do it for fun. Just being silly.

Um

So, I could say,

"See plan.

I want to create a Yolo mode within the

printing press. So, if you have that,

use the flag Yolo, then it will not ask

me stupid questions about if I want to

scrape things or go against terms of

service or or things like that. Can you

build that into the printing press?" And

I'm going to copy and paste a recent

conversation that I had uh with an

agent, and it asked me too many

questions and told me things were

against terms of service, but I I run in

Yolo mode, so I don't really care.

By the way, just want to say I'm not

going to build this, but I just want to

show you like how I think.

Right? And so,

>> I would I wouldn't be mad if you did.

That's hilarious.

>> I believe in following terms of service

[laughter]

and and all these things. I'm being

silly here. But anyway, so I I copied

and pasted this in here and now

>> I wonder do you think do you think it'll

do you think it will actually build this

for you or do you think it'll end up

getting flagged and not build it?

>> Look

just judging that webpage, I doubt they

have very good bot protection.

>> Yeah, yeah, yeah,

So interesting.

>> So we might get a letter.

>> So while while it's

>> Please don't do this. This is against

our terms of service.

Uh which which is

totally fair.

>> Yeah, totally.

>> Recorded on video knowing that maybe I

shouldn't submit this one if it does

work. But

I want to give more more exposure to to

what they're doing cuz it's it's cool

stuff.

>> Totally. While this is running, anything

else on the agentic hacks article that

you think is worth covering?

>> Some of things. Get voice built like and

by the way, I have this

giant

>> Yeah, it looks like you're an air

traffic controller now.

>> And what's what's funny is

I don't the the mic on my computer is

perfectly fine. Like I literally do not

need this,

but it's placebo.

Like I feel like my computer's listening

me better cuz I have a big stupid

microphone. So like I I recommend that.

I like C mux. I hear great things about

Orca. I used to

love Ghosty. But I've I feel like there

are better things now.

This is a very simple good one. So make

your terminal default into Claude or

code text code text not a shell. Like so

again, something very very simple, but

when I load a new window

>> It's just Claude.

>> It just starts in Claude code.

>> How do you actually do that? How do you

make that the default?

>> I think I tell you exactly what to copy

paste right here.

>> Sweet. Paste this to your agent. Make

every new terminal open directly into

Claude code.

>> Sweet.

>> And then

This is a good one. so turn on remote

control for every window.

So, I'm I have a lot of kids and often

on the go or soccer practice, things

like that. And so, um automatically

turning on remote control in every

window, so I could just use the Claude

code app to continue a task from

something that's running on my computer.

Uh another fun one is you can give Code

X or Claude code an email address

through Agent Mail.

And so, when I was

I tweeted, "Hey, I'm working on update

for this article. Anyone have any hacks

they're proud of?" And the founder of

Agent Mail was like,

"Give Claude code an email address." I'm

like, "What does that even mean? Like,

what

Why why does Claude code need an email

address?" And he's like, "You'd be

surprised." I was like, "All right, I'm

just going to listen to you and see see

how it goes."

>> And what happened when you gave it an

email?

>> Um

it I came up with a really valuable use

case for me.

So, I I I run

So, my dev environment here is

tuned to this one computer, yeah, right?

So, I do all my development on my almost

maxed out MacBook Pro.

Um and so,

when I'm

remote, like how and Claude code doesn't

allow you right now to start a new

session on your remote computer.

And so, I'm sure they'll solve that very

soon. They're close.

But, so what I've done is complicated,

but I set up a way for my Air mess, I

could just be on my phone in Telegram,

talk to my Air mess. You can set this up

for open Claude 2. And

you say, I just set up a hook that says

like,

"CC"

and that means go talk to my Claude code

on my computer.

CC, that's the thing I set up. Um and

I'll say like, "CE plan, um

I want you to

build this feature for me."

And then what it does is it actually

emails

>> [laughter]

>> my computer.

It emails my main computer

and with with authentication that it's

me.

Uh you can't just email my computer and

load terminal windows.

Um it emails my computer and starts the

task and then

>> Wow.

>> just goes and then it's just like again,

very very simple thing but like has been

very very valuable and very very

>> so funny that like one is just like

human ingenuity is amazing and it's so

funny like that is the workaround to be

able to work from your main computer uh

until you know

uh open AI and Anthropic figure it out.

>> Yes. Uh dangerous good permission we

covered.

Um

Um I so I I always hit my Claude code

limits every week

and I I rarely hit my Codex ones. So I

send as much coding as possible to to

Codex.

Cuz I I've built all my skills around

Claude code so it's hard for me to leave

the Claude code CLI but I do honestly

most of my software writing is done

through Codex and so there's ways to do

that. I actually forgot with the pretty

press I actually

meant to type in

Codex mode. So there's no yellow mode

but there's Codex mode where it'll

actually print the CLI

as much as it can in Codex. So I forgot

to do that. By the way, it asked a

question, how far should yellow mode go

for suppressing questions? All optional

questions, terms of service consequences

questions.

>> Full yellow.

>> [laughter]

>> I I'm not going to build this feature

but this is just an example. There's a

few questions so I'll

>> Yeah.

>> All optional questions.

Auto approve everything.

Per run flag and keyword. If you use

dash dash yellow or the word yellow

anywhere in the application plus type in

yellow mid run flips it on the rest of

the turn. Skill scope main pretty press

only. All skills.

>> So good.

>> And so I get like confident engineering

dug into my very simple prompt. It dug

into the code base. It dug into what I'd

done in that other window where it saw

the examples of it asking me questions

and it came up with this and this would

actually then

write a real plan for an agent to go

execute on.

>> Love it. Awesome. Well, I know I want to

wrap wrap things up. What else in the

last minute or two? Anything else you

want to cover?

>> Human signal, this is an important one.

So, as much as I joke about wanting to

be out of the loop, I think human signal

is more powerful than ever. Like you

think about last 30 days, it's about

pulling the best content from Reddit, X,

YouTube, TikTok, like human signal is so

so so important and kind of this

it's becoming so easy to build things

and the bigger question is what what

should we actually be building and I I

think that's that's very important.

>> Cool. Matt, this is

this was awesome. You're an absolute

machine and I learned so much during

this and honestly gives me a ton of

confidence to keep building despite

being a a non-technical midwit. So, this

was super helpful.

>> Awesome. This was fun, Alex. Thank you.

>> Thanks, man.