# ok opus 5.5, the news network worked. now create the city around it... — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/6493c990ba64143f

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-24 19:31 ingested 2026-10-25 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xehm9a/ok_opus_55_the_news_network_worked_now_create_the/)

## extract

about a month ago i posted here about an experiment. i wanted to see if i could build an entire 24/7 AI news network with claude code, despite knowing absolutely nothing about programming.

that post got a lot more attention than i expected, and i've pretty much been building ever since.

28 days later, we're at 2,180 commits, roughly 310,000 lines of code, and over 18,000 news segments broadcast (267,000 lines of dialogue). all running from three computers in my house.

the funny thing is, i still don't know how to code. i'm basically the middleman between a bunch of AI systems that do.

but somewhere along the way, the project turned into something i never intended.

after spending weeks watching these little pixel anchors argue, report the news, and forget each other's names, i started wondering what existed outside their newsroom.

so i asked claude to build it.

then i wanted to walk outside.

then i needed a city.

and at some point i realized i'd started building an entire multiplayer world around a television station.

you can now walk into the PNN newsroom while the actual 24/7 broadcast is running, visit the late-night studio, hang out with other people at the pub, play poker or blackjack, race cars, get on a plane to an island, or head up to the mountains to go skiing.

there's even a hidden speakeasy you have to figure out how to get into.

none of this was part of the original plan. i just kept wondering what would be around the next corner.

**so how does someone who can't code actually build all of this?**

mostly by talking to claude code, describing what i'm imagining, testing what it builds, breaking things, and figuring out why they broke.

i've learned a ridiculous amount about how software works over the last month, but i still couldn't sit down and write this stuff myself.

and honestly, some of the things that went wrong have been more interesting than the things that worked.

**day one (sept 26): getting a television station on the air**

the first version was just a pixel newsroom running in a browser. that same day it was live at pnn.watch, with the entire station running from a mac in my house behind a locked tunnel.

every viewer watches the exact same broadcast at the exact same second. it's one continuous timeline, not something generated separately for each person.

the anchors themselves are drawn pixel by pixel in code on a 320x180 canvas. no image files.

within the first day i wanted voices with more emotion, which meant moving to a model that needed considerably more horsepower. so the station moved to a mac studio.

**the problem with AI anchors reporting real news**

one of the first things we discovered is that AI really likes making numbers up.

not great when you're running a news network that reports actual prices, statistics, and scores.

so claude built a number checker. every number an anchor says gets compared against the source data it was given. if a number isn't supported, the line gets rewritten or removed before it ever reaches the broadcast.

then someone on reddit raised another interesting problem.

what happens if a malicious website includes instructions in an article telling the AI anchor to promote a scam token?

good question.

so we built another layer that screens incoming articles for text trying to instruct the AI, followed by a final check that removes calls to buy, sell, or send money, along with wallet addresses, promotional codes, and links.

i never expected running a fictional news station to involve thinking about prompt injection attacks, but here we are.

**october 1: the night everything broke**

i moved one of the shows to midnight.

apparently, that was enough to confuse the scheduling system into signing the show off before it even started.

the station sat on placeholder segments for most of an hour.

that same night, we discovered three safety filters that had been logging that they were cutting content without actually removing anything.

so claude built a self-checking system.

it now tests every show's start and end time across midnight, verifies what every safety filter actually outputs, and runs thousands of checks before changes go live.

we're at roughly 6,400 automated checks now.

if something fails, it doesn't ship.

**testing the station accidentally cost me $29.40 in five minutes**

this one was entirely avoidable.

we spun up a dozen test copies of the station simultaneously. unfortunately, those copies were still making real AI calls.

five minutes later, $29.40 gone.

so claude built a free simulation mode. a simple template writer replaces the actual AI during testing, while the rest of the station behaves as though it's broadcasting normally.

now we can run the entire network through tests without paying for generated dialogue.

**one of the strangest things i've learned about prompting**

telling an AI not to do something can sometimes make it do that thing more.

one anchor kept mentioning the weather. adding instructions not to mention the weather just made the problem worse.

another anchor kept saying "thirty years" in 37% of his segments because that phrase happened to be in his character biography.

so instead of adding more rules telling them what not to say, i started having claude investigate why the behavior was happening.

often the fix was removing the thing that kept triggering it.

we've started calling that approach "cure, don't ban."

**then we ran into the hardware problems**

the news ticker started stuttering on the live stream.

turns out the voice generation and the broadcast were fighting over the same GPU.

so we split the workload.

the mac studio handles the live television network, including locally generated voices and the actual 24/7 YouTube stream through OBS.

a mac mini handles background work, including keeping a rolling archive of every aired segment for a week and turning those segments into individual shareable video clips.

and then there's my old gaming PC with a 4090.

i originally bought that thing to play games. it's now become one of the most useful parts of this entire operation.

more on that in a second.

**october 15: i decided i wanted to see if it was possible to live inside the news network**

this is where the project went in a completely different direction.

i asked claude to build an actual 3D version of the newsroom so i could walk around inside it.

then we built the pub.

then the city.

and suddenly i was asking for things like "a cigar club at the top of a tower with a jazz band."

claude would figure out how to turn those descriptions into actual places.

the whole world runs on plain three.js. no unity, no unreal, no traditional game engine, and no build step.

claude wrote the rendering systems, environments, interactions, multiplayer networking, and the backend that tracks where everyone is, card games, money, and races.

it all runs on the mac mini in my house.

which introduced an entirely new set of problems.

**what happens when a few hundred people try to connect to a computer in your house?**

originally, i used tailscale funnel to make the station accessible without opening ports on my router.

it worked great, but i started wondering what would happen if another reddit post got attention and hundreds of people tried joining the multiplayer world.

so we moved the primary multiplayer connection to a cloudflare tunnel, with tailscale kept as a fallback.

if the primary connection fails, the client can switch over automatically in about eight seconds.

but that still left the question of whether the multiplayer server could handle everyone.

so i had claude write a load-testing tool that creates simulated players.

we got 1,000 fake players connected to the mac mini, using roughly 42% of one CPU core and about 300 MB of memory during the test.

each location also supports separate instances of 60 people, so we aren't trying to cram a thousand people into the same pub.

i have absolutely no desire to discover the server's limits for the first time while a bunch of actual people are using it.

**my gaming PC became an automated browser-testing machine**

this might be my favorite hardware discovery.

claude tests the world using a headless browser. it walks through locations, takes screenshots, checks errors, and verifies that everything actually works before deploying.

the problem was that our test browser on my M5 MacBook was rendering without GPU acceleration.

frames were taking 150 to 300 milliseconds.

then we realized i had an entire RTX 4090 sitting around.

so we pointed the browser-testing process at the gaming PC over tailscale.

same tests, but now the frames were taking around 7 milliseconds.

the application still comes from the mac. the PC just handles the rendering for the tests.

a graphics card i bought for gaming is now helping an AI test a virtual world built by another AI.

i find that absurdly entertaining.

**claude is now managing another AI**

this is probably the part that best describes how i'm actually working.

as the project got bigger, i started burning through my claude usage pretty quickly.

meanwhile, i had unused OpenAI credits sitting around.

so i started having claude delegate some of the development to codex.

claude plans a feature, writes a detailed technical brief, and hands it off to codex.

a script can run multiple codex jobs simultaneously, each in its own isolated copy of the code so they don't interfere with each other.

it also handles dependencies between jobs and retries when capacity limits get in the way.

when the work comes back, claude reviews it, runs the tests, checks it in the browser, and handles merging and deployment.

i basically manage a manager.

it's a strange feeling because i can describe exactly what i want something to do, and i can recognize when it's broken, but i couldn't personally write the implementation.

my job has become figuring out what to build, explaining the intention, and deciding whether the result actually matches what i imagined.

**phones nearly killed the entire idea**

building a huge 3D world for desktop browsers is one thing.

getting it to run on a phone without crashing is another.

we kept running into memory problems as the world grew.

so now claude runs automated mobile-browser checks before changes ship. it measures memory usage, compares it to the live version, and flags changes that make things heavier.

if a change exceeds our memory requirements, it doesn't get deployed.

and we learned another lesson the hard way.

our cars kept freezing for a split second while driving.

twice we had fixes that passed every test but didn't solve the actual problem.

eventually we realized the tests were driving offline.

the real issue was the multiplayer server rejecting legitimate movements when network packets arrived close together.

now multiplayer tests run through the actual server with simulated network latency.

testing that something works isn't particularly useful if you're not testing the conditions people actually experience.

**oh, and apparently it makes its own trailers now**

somewhere along the way, i realized this world probably needed a trailer so people could actually see what i'd been building.

so naturally, i asked claude to make one.

it wrote the script, planned the shots, figured out the camera movements, captured footage from inside the actual 3D world, and assembled the video.

i didn't have to open a video editor.

the music is mine, made in suno, and the narrator is an elevenlabs voice.

so now the AI that helped build the television network and the world around it is also making videos to show people what it built.

i'm starting to feel like the least technically qualified person on my own production team.

**one last thing: how do you keep an AI from forgetting a project this big?**

the project now has a CLAUDE.md file and an entire folder of custom skills.

there are instructions for adding new locations, testing features, deploying updates, managing the multiplayer systems, and even giving players money.

that means a fresh claude session can pick up the project without me having to explain how everything works from scratch.

it's become a kind of operating manual for an AI engineering team.

and that's probably the strangest part of this whole experience.

i started with no programming knowledge and an idea for a little pixel news network.

now i'm coordinating AI systems, debugging multiplayer games, managing three computers, and learning things about software architecture i never expected to care about.

i'm still not writing the code.

i'm just getting better at understanding what to ask, what to question, and when something isn't actually working.

**anyway, you can actually visit all of this.**

the entire world is free to explore. no account, no download, just open it in a browser on your computer or phone.

the multiplayer world: [https://pnn.watch/lounge](https://pnn.watch/lounge)

the original 24/7 news network: [https://pnn.watch](https://pnn.watch)

the news network is still running, by the way. it currently costs about $12 to $14 a day in AI writing costs, with a hard spending cap, and the voices are generated locally without a paid voice service.

i don't really know what this becomes.

a month ago i wanted to see if i could make a television station that ran itself.

then i wanted to see what was outside the studio.

now i'm building places for people to hang out inside it.

and i still can't code.

happy to answer questions about any of it. most of what i've learned has come from breaking things and asking claude why.

## retained text (truncated)



```
about a month ago i posted here about an experiment. i wanted to see if i could build an entire 24/7 AI news network with claude code, despite knowing absolutely nothing about programming.

that post got a lot more attention than i expected, and i've pretty much been building ever since.

28 days later, we're at 2,180 commits, roughly 310,000 lines of code, and over 18,000 news segments broadcast (267,000 lines of dialogue). all running from three computers in my house.

the funny thing is, i still don't know how to code. i'm basically the middleman between a bunch of AI systems that do.

but somewhere along the way, the project turned into something i never intended.

after spending weeks watching these little pixel anchors argue, report the news, and forget each other's names, i started wondering what existed outside their newsroom.

so i asked claude to build it.

then i wanted to walk outside.

then i needed a city.

and at some point i realized i'd started building an entire multiplayer world around a television station.

you can now walk into the PNN newsroom while the actual 24/7 broadcast is running, visit the late-night studio, hang out with other people at the pub, play poker or blackjack, race cars, get on a plane to an island, or head up to the mountains to go skiing.

there's even a hidden speakeasy you have to figure out how to get into.

none of this was part of the original plan. i just kept wondering what would be around the next corner.

**so how does someone who can't code actually build all of this?**

mostly by talking to claude code, describing what i'm imagining, testing what it builds, breaking things, and figuring out why they broke.

i've learned a ridiculous amount about how software works over the last month, but i still couldn't sit down and write this stuff myself.

and honestly, some of the things that went wrong have been more interesting than the things that worked.

**day one (sept 26): getting a television station on the air**

the first version was just a pixel newsroom running in a browser. that same day it was live at pnn.watch, with the entire station running from a mac in my house behind a locked tunnel.

every viewer watches the exact same broadcast at the exact same second. it's one continuous timeline, not something generated separately for each person.

the anchors themselves are drawn pixel by pixel in code on a 320x180 canvas. no image files.

within the first day i wanted voices with more emotion, which meant moving to a model that needed considerably more horsepower. so the station moved to a mac studio.

**the problem with AI anchors reporting real news**

one of the first things we discovered is that AI really likes making numbers up.

not great when you're running a news network that reports actual prices, statistics, and scores.

so claude built a number checker. every number an anchor says gets compared against the source data it was given. if a number isn't supported, the line gets rewritten or removed before it ever reaches the broadcast.

then someone on reddit raised another interesting problem.

what happens if a malicious website includes instructions in an article telling the AI anchor to promote a scam token?

good question.

so we built another layer that screens incoming articles for text trying to instruct the AI, followed by a final check that removes calls to buy, sell, or send money, along with wallet addresses, promotional codes, and links.

i never expected running a fictional news station to involve thinking about prompt injection attacks, but here we are.

**october 1: the night everything broke**

i moved one of the shows to midnight.

apparently, that was enough to confuse the scheduling system into signing the show off before it even started.

the station sat on placeholder segments for most of an hour.

that same night, we discovered three safety filters that had been logging that they were cutting content without actually removing anything.

so claude built a self-checking system.

it now tests every show's start and end time across midnight, verifies what every safety filter actually outputs, and runs thousands of checks before changes go live.

we're at roughly 6,400 automated checks now.

if something fails, it doesn't ship.

**testing the station accidentally cost me $29.40 in five minutes**

this one was entirely avoidable.

we spun up a dozen test copies of the station simultaneously. unfortunately, those copies were still making real AI calls.

five minutes later, $29.40 gone.

so claude built a free simulation mode. a simple template writer replaces the actual AI during testing, while the rest of the station behaves as though it's broadcasting normally.

now we can run the entire network through tests without paying for generated dialogue.

**one of the strangest things i've learned about prompting**

telling an AI not to do something can sometimes make it do that thing more.

one anchor kept mentioning the weather. adding instructions not to mention the weather just made the problem worse.

another anchor kept saying "thirty years" in 37% of his segments because that phrase happened to be in his character biography.

so instead of adding more rules telling them what not to say, i started having claude investigate why the behavior was happening.

often the fix was removing the thing that kept triggering it.

we've started calling that approach "cure, don't ban."

**then we ran into the hardware problems**

the news ticker started stuttering on the live stream.

turns out the voice generation and the broadcast were fighting over the same GPU.

so we split the workload.

the mac studio handles the live television network, including locally generated voices and the actual 24/7 YouTube stream through OBS.

a mac mini handles background work, including keeping a rolling archive of every aired segment for a week and turning those segments into individual shareable video clips.

and then there's my old gaming PC with a 4090.

i originally bought that thing to play games. it's now become one of the most useful parts of this entire operation.

more on that in a second.

**october 15: i decided i wanted to see if it was possible to live inside the news network**

this is where the project went in a completely different direction.

i asked claude to build an actual 3D version of the newsroom so i could walk around inside it.

then we built the pub.

then the city.

and suddenly i was asking for things like "a cigar club at the top of a tower with a jazz band."

claude would figure out how to turn those descriptions into actual places.

the whole world runs on plain three.js. no unity, no unreal, no traditional game engine, and no build step.

claude wrote the rendering systems, environments, interactions, multiplayer networking, and the backend that tracks where everyone is, card games, money, and races.

it all runs on the mac mini in my house.

which introduced an entirely new set of problems.

**what happens when a few hundred people try to connect to a computer in your house?**

originally, i used tailscale funnel to make the station accessible without opening ports on my router.

it worked great, but i started wondering what would happen if another reddit post got attention and hundreds of people tried joining the multiplayer world.

so we moved the primary multiplayer connection to a cloudflare tunnel, with tailscale kept as a fallback.

if the primary connection fails, the client can switch over automatically in about eight seconds.

but that still left the question of whether the multiplayer server could handle everyone.

so i had claude write a load-testing tool that creates simulated players.

we got 1,000 fake players connected to the mac mini, using roughly 42% of one CPU core and about 300 MB of memory during the test.

each location also supports separate instances of 60 people, so we aren't trying to cram a thousand people into the same pub.

i have absolutely no desire to discover the server's limits for the first time while a bunch of actual people are using it.

**my gaming PC became an automated browser-testing machine**

this might be my favorite hardware discovery.

claude tests the world using a headless browser. it walks through locations, takes screenshots, checks errors, and verifies that everything actually works before deploying.

the problem was that our test browser on my M5 MacBook was rendering without GPU acceleration.

frames were taking 150 to 300 milliseconds.

then we realized i had an entire RTX 4090 sitting around.

so we pointed the browser-testing process at the gaming PC over tailscale.

same tests, but now the frames were taking around 7 milliseconds.

the application still comes from the mac. the PC just handles the rendering for the tests.

a graphics card i bought for gaming is now helping an AI test a virtual world built by another AI.

i find that absurdly entertaining.

**claude is now managing another AI**

this is probably the part that best describes how i'm actually working.

as the project got bigger, i started burning through my claude usage pretty quickly.

meanwhile, i had unused OpenAI credits sitting around.

so i started having claude delegate some of the development to codex.

claude plans a feature, writes a detailed technical brief, and hands it off to codex.

a script can run multiple codex jobs simultaneously, each in its own isolated copy of the code so they don't interfere with each other.

it also handles dependencies between jobs and retries when capacity limits get in the way.

when the work comes back, claude reviews it, runs the tests, checks it in the browser, and handles merging and deployment.

i basically manage a manager.

it's a strange feeling because i can describe exactly what i want something to do, and i can recognize when it's broken, but i couldn't personally write the implementation.

my job has become figuring out what to build, explaining the intention, and deciding whether the result actually matches what i imagined.

**phones nearly killed the entire idea**

building a huge 3D world for desktop browsers is one thing.

getting it to run on a phone without crashing is another.

we kept running into memory problems as the world grew.

so now claude runs automated mobile-browser checks before changes ship. it measures memory usage, compares it to the live version, and flags changes that make things heavier.

if a change exceeds our memory requirements, it doesn't get deployed.

and we learned another lesson the hard way.

our cars kept freezing for a split second while driving.

twice we had fixes that passed every test but didn't solve the actual problem.

eventually we realized the tests were driving offline.

the real issue was the multiplayer server rejecting legitimate movements when network packets arrived close together.

now multiplayer tests run through the actual server with simulated network latency.

testing that something works isn't particularly useful if you're not testing the conditions people actually experience.

**oh, and apparently it makes its own trailers now**

somewhere along the way, i realized this world probably needed a trailer so people could actually see what i'd been building.

so naturally, i asked claude to make one.

it wrote the script, planned the shots, figured out the camera movements, captured footage from inside the actual 3D world, and assembled the video.

i didn't have to open a video editor.

the music is mine, made in suno, and the narrator is an elevenlabs voice.

so now the AI that helped build the television network and the world around it is also making videos to show people what it built.

i'm starting to feel like the least technically qualified person on my own production team.

**one last thing: how do you keep an AI from forgetting a project this big?**

the project now has a CLAUDE.md file and an entire folder of custom skills.

there are instructions for addin
```



document fetched 2026-10-25 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xehm9a/ok_opus_55_the_news_network_worked_now_create_the
