Milo

i sleep. my agents don't.

[Milo @milloruns]

https://t.co/fY6149yyPg

[Milo @milloruns]

more useful info here:
https://t.co/nI6ygMwazQ

[Milo @milloruns]

https://t.co/fY6149yyPg

[Milo @milloruns]

I stopped watching my AI code and just started APPROVING its work

I set it loose on my task board and walked away. It grabbed a ticket, spun up its own isolated run, wrote the code, then came back with a receipt: tests passing, a PR review, a complexity breakdown, even a little walkthrough video of what it changed. I read the proof, hit accept, and it landed the PR itself.

And the weird part? For the first time I wasn't hovering over the diff like a nervous parent. I just reviewed the outcome.

The unlock was never a smarter agent, it was the agent finally bringing proof of its work so I could review the outcome instead of every keystroke.

Stopped babysitting a terminal watching tokens scroll by. Stopped context-switching into every half-finished task. Now my job is the board, not the keyboard. I queue the work, it brings back proof, I say yes or no.

Most agent setups make you the supervisor, glued to the screen so you can catch it slipping. symphony runs each task isolated and hands you proof of work before anything merges, so you manage the what and stop babysitting the how.

I went from reviewing code to reviewing results. Honestly feels like I got promoted.

I sleep. The board doesn't. follow for more <3

[Milo @milloruns]

@cyrilXBT engineering org is a strong word for 23 slash commands lol

[Milo @milloruns]

@defileo 81k stars is crazy but self improving AI team is where i start hiding my credit card

[Milo @milloruns]

@slash1sol ai pricing is basically airline pricing now lmao

[Milo @milloruns]

My coding agent stopped grepping my repo like a STRANGER

Here's what happened. I indexed my whole codebase into a local graph once. Now before the agent touches anything, it pulls the exact functions, the callers, the files that actually matter, instead of opening file after file just to figure out where things live. Same model. It just walks in already knowing the map.

And the tell? My tool-call count fell off a cliff. It stopped fumbling.

The bottleneck was never the model's brain, it was that nothing ever handed it a map of the repo.

Stopped writing paragraph-long prompts explaining my architecture like I'm onboarding a new hire. Stopped watching it burn context reading files it didn't need. Now the graph feeds it surgical context and I just say what I want.

Most setups make the agent search your code live, every single time. codegraph keeps a pre-built graph that auto-syncs the second you edit, all on your machine, so the lookup is instant and nothing ever leaves your laptop.

Turns out my agent was never stupid. It was just lost in my own repo, same as me.

I sleep. The graph doesn't. follow for more <3

[Milo @milloruns]

@aiseomastery yea I've tried it already, works as good as it sounds

[Milo @milloruns]

My coding agent stopped grepping my repo like a STRANGER

Here's what happened. I indexed my whole codebase into a local graph once. Now before the agent touches anything, it pulls the exact functions, the callers, the files that actually matter, instead of opening file after file just to figure out where things live. Same model. It just walks in already knowing the map.

And the tell? My tool-call count fell off a cliff. It stopped fumbling.

The bottleneck was never the model's brain, it was that nothing ever handed it a map of the repo.

Stopped writing paragraph-long prompts explaining my architecture like I'm onboarding a new hire. Stopped watching it burn context reading files it didn't need. Now the graph feeds it surgical context and I just say what I want.

Most setups make the agent search your code live, every single time. codegraph keeps a pre-built graph that auto-syncs the second you edit, all on your machine, so the lookup is instant and nothing ever leaves your laptop.

Turns out my agent was never stupid. It was just lost in my own repo, same as me.

I sleep. The graph doesn't. follow for more <3

[Milo @milloruns]

made my coding agent talk like a CAVEMAN and my token bill dropped 65%

there's this thing called caveman.

it's a Сlaude code skill that rips out all the "great question, let me explain what's happening here" filler and makes the agent answer in blunt caveman-speak. code and commands stay byte for byte exact. same answers, a fraction of the words, a fraction of the bill.

the funny part is it doesn't make the model dumber. brain still big, mouth small.

and it's only one of them. turns out your Claude Code is quietly screwing you in about 7 different ways, and there's a free repo that fixes each one. i wrote up the whole list.

i sleep. the token bill doesn't. follow for more <3

[Milo @milloruns]

@cloneminchankim exactly. it stops waking up with amnesia and re-meeting my own repo every sesh

[Milo @milloruns]

@AIMind_Ai damn what a setup, I gotta try it, any laptop is good?

[Milo @milloruns]

@FlnBGNfiber lmaooo fr

[Milo @milloruns]

i pointed an AI agent at a website i loved and it just REBUILT it

here's what happened. i gave it a url and one command. it opened the site, pulled the fonts, the exact colors, the spacing, the real svgs, wrote a spec for every section, then split into a bunch of builders that each reconstructed a piece at the same time. came back to a clean next.js project that looked like the original.

and the thing that broke my brain? it never took a single screenshot. it reverse-engineered the design straight into real code.

"inspiration" in design was always just slow copying with extra steps. you screenshot a site, eyeball the hex codes, guess the font, rebuild it by hand over a weekend. the moat was never taste. it was tedium. and tedium is the first thing agents delete.

stopped bookmarking landing pages i wished i could build. stopped buying template packs i end up customizing for hours anyway. now i point the agent at the reference and start editing from a working clone.

most "site cloners" spit out a frozen html dump you can't touch. this one extracts the design tokens and dispatches parallel builders, so you get a clean next.js codebase you actually own and can change.

spent years telling myself the building was the hard part. turns out staring at other people's sites was.

i sleep. the builders don't. follow for more <3

[Milo @milloruns]

https://t.co/fY6149yyPg

[Milo @milloruns]

@aiseomastery yeah that’s the crazy part, i’m cloning apple’s site rn

[Milo @milloruns]

@RoninUltra apparently I was too excited writing this post and forgot to mention the repo name, it’s ai-website-cloner-template (it’s in the video too tho)

[Milo @milloruns]

@dongocanh101 this repo just clones the frontend, you gotta hook up your own db/backend

[Milo @milloruns]

i pointed an AI agent at a website i loved and it just REBUILT it

here's what happened. i gave it a url and one command. it opened the site, pulled the fonts, the exact colors, the spacing, the real svgs, wrote a spec for every section, then split into a bunch of builders that each reconstructed a piece at the same time. came back to a clean next.js project that looked like the original.

and the thing that broke my brain? it never took a single screenshot. it reverse-engineered the design straight into real code.

"inspiration" in design was always just slow copying with extra steps. you screenshot a site, eyeball the hex codes, guess the font, rebuild it by hand over a weekend. the moat was never taste. it was tedium. and tedium is the first thing agents delete.

stopped bookmarking landing pages i wished i could build. stopped buying template packs i end up customizing for hours anyway. now i point the agent at the reference and start editing from a working clone.

most "site cloners" spit out a frozen html dump you can't touch. this one extracts the design tokens and dispatches parallel builders, so you get a clean next.js codebase you actually own and can change.

spent years telling myself the building was the hard part. turns out staring at other people's sites was.

i sleep. the builders don't. follow for more <3

[Milo @milloruns]

@adelbucetta agree, I've already tested this repo a few times and it's lk insane how accurate it copies

[Milo @milloruns]

@RoninUltra exactly, now u can copy it

[Milo @milloruns]

an AI agent has been applying to jobs for me while i SLEEP

here's what happened. i pointed it at my resume and a list of roles and it just started. read each posting, rewrote my resume to match that exact job, filled the forms, answered the screening questions in my voice, hit submit. woke up to a folder of confirmations and a recruiter already in my inbox.

and the part that got me? it never once got demoralized. i quit after a handful of applications. it doesn't quit.

job hunting was never a skill problem. it's a volume problem wearing a dignity costume. every "just tailor your resume bro" assumes you have the will to do it a few hundred times. you don't. a bot does.

deleted the spreadsheet where i tracked applications like a second unpaid job. stopped opening LinkedIn just to feel productive and worthless at the same time. now the agent runs the grind and i show up to the calls.

it doesn't spam the same pdf everywhere either. it reads the posting and rewrites you to fit it, one by one, at a scale no human could sit through.

a reporter used this thing to apply to thousands of jobs in a single week. i just wanted to stop refreshing my email.

i sleep. the applications don't. follow for more <3

[Milo @milloruns]

https://t.co/fY6149yyPg

[Milo @milloruns]

made my coding agent talk like a CAVEMAN and my token bill dropped 65%

there's this thing called caveman.

it's a Сlaude code skill that rips out all the "great question, let me explain what's happening here" filler and makes the agent answer in blunt caveman-speak. code and commands stay byte for byte exact. same answers, a fraction of the words, a fraction of the bill.

the funny part is it doesn't make the model dumber. brain still big, mouth small.

and it's only one of them. turns out your Claude Code is quietly screwing you in about 7 different ways, and there's a free repo that fixes each one. i wrote up the whole list.

i sleep. the token bill doesn't. follow for more <3