Pinned Tweet
I'm Ruslan. I've spent 17 years writing backend code, mostly PHP, high-load systems and financial integrations. The kind of stuff where you really don't want to screw things up))
I use Codex and Claude a lot now. Getting them to write code is usually the easy part. Then I have to check what they actually did 😅
Over the last few years, I've tried launching several projects. I got most of an AI coding school built before getting stuck on finding students. A transcription app made a few sales, and I closed it anyway.
Right now I'm focused on Nourmio, a calorie and sugar tracking app we built with friends.
I'll share what we're building and how we try to get people to use it. There'll probably be other projects along the way too.
If you're building your own thing, say hi. I'd like to hear what you're working on.
sugarwise.monster/
2
23
1,378
Paying for Claude apparently comes with a side quest: checking whether your internet connection looks suspicious 😅
This checker keeps coming up in ban threads:
ip.net.coffee/claude/
It shows IP location, hosting/VPN/proxy flags, DNS/WebRTC leaks and timezone mismatches. Useful for spotting odd network settings.
But I couldn't verify that its ‘Trust Score’ matches Anthropic's own checks. A green result doesn't tell you why an account was banned. And if Claude Code runs on a remote server, testing your laptop's browser isn't testing that server.
I went through the ban threads and official docs. Here's what I'd actually try:
1. Read the exact error first.
A broken login isn't the same as a suspension. Claude had an authentication incident today, Nov 13. Check status.claude.com before changing half your setup.
For login errors, Anthropic suggests turning off VPN, disabling extensions and clearing cache/cookies. That may help a login problem. It won't reverse a ban or make an unsupported region eligible.
2. Check the actual account rules.
Anthropic lists repeated Usage Policy violations, account creation from unsupported locations and Terms violations as reasons for bans. An IP checker can't tell you whether your account meets all those rules.
3. If it's identity verification, use the official flow.
Persona asks for an original valid photo ID, sometimes a live selfie. Passing it doesn't guarantee access. If a photo is unreadable, the official advice is to retry with better lighting and a legible document. Buying a ‘verified’ account isn't the same thing.
4. If you're actually suspended, appeal.
Log into the banned account at claude.ai and use the appeal form. I'd include the exact message and what happened before it. Check the export option too, although available data can be restricted.
One Reddit user reported getting access back in five hours. Anthropic currently warns of longer appeal response times. Nobody can promise you a deadline.
I'd be very careful with anyone selling a guaranteed unban or a magic score threshold.
If you got reinstated, did support explain the trigger, or just restore access?
Official pages behind the checklist:
Login errors:
support.claude.com/en/articl…
Identity verification:
support.claude.com/en/articl…
Appeals and data export:
support.claude.com/en/articl…
The five-hour recovery is one user's report, not a promised response time:
reddit.com/r/ClaudeCode/comm…
1
1
198
The PR demo I want includes the bit that goes wrong 😅
For a food-logging change in Nourmio, I'd want to see a wrong guess, someone correcting the dish, recalculating, then reopening the saved meal.
If the new numbers look right on screen but the diary still has the old calories, I want to find that out before merging.
An animation can explain the change. I'd also attach a recording from the actual build and say which commit it came from.
Show me the taps, the waiting and what got saved. That would make the review a lot easier.
My dream is for every important PR to be accompanied by a video that shows folks how it works
Working on an /animatic skill for this
2
1
99
The $0 output-token bill caught my eye here 😅
Google's new Decision 1 scores fixed choices at $0.10 per million input tokens.
For Nourmio, I'd try it before food recognition: use this photo, ask for a clearer one, or ask what's in the dish.
I haven't tested it yet. I'd want to see how many bad estimates that extra step prevents, and how many people give up logging their meal because we asked another question.
People are trying to log lunch. Every extra question had better earn its place.
introducing the Gemini Decisions API
powered by Gemini Decision 1, our new multimodal decision model
get fast, calibrated decisions on up to 1000 questions at once across text, images, audio, video, and PDFs
try it today in AI Studio and via the Gemini API ai.studio
1
2
105
Someone tried Nourmio on an iPhone 13 and sent me the kind of feedback I can actually work with.
They said the keyboard was covering the voice recording button. And “Enter your portion to see the total” looked like a field they were supposed to type into.
Both are now on my list to reproduce. One gets in the way of recording a meal. The other makes someone hunt for where to put a number 😅
When you build the app, you already know where everything is meant to be. Someone opening it for the first time doesn’t have that cheat sheet.
If you try Nourmio, tell me the first thing that makes you stop and think, “where the hell do I tap?”
A screenshot and your phone model give me somewhere to start.
3
101
“Banked reset” is the bit of this Codex tease I’m watching 😅
Tibo is hinting at something for Thursday, Nov 12. Still a hint, so let’s see what actually ships.
I want enough quota to finish an annoying bug properly. Read the repo, try a fix, find out why it’s wrong, run the tests again. That whole loop counts as one task in my head))
A more capable model gets a lot less exciting if you’ve used the week’s allowance figuring out where the bug lives.
What eats your Codex limit faster: building new features or chasing bugs?
Replying to @thsottiaux
Thursday is feeling codexy to me. Perhaps also a good time for a banked reset.
2
2
189
If ChatGPT recommends Nourmio 1,000 times, I want to know what happens after the click.
Juozas Kaziukėnas spotted a test screen for mentions and clicks. I’d add one number on my end. How many of those visitors tap through to the App Store?
linkedin.com/posts/juozas_ch…
2
5
254
Give the next agent a way to check the note.
“Be careful with payments” will just sit there in CLAUDE.md. “Replay this webhook twice and check that there’s still one order” gives it something to run.
My takeaway from the 2025 ACE paper. arxiv.org/abs/2510.04618
holy f*ck... the best AI agents don't have a better model. they have better notes
Stanford University just published the whole method for free in a paper called Agentic Context Engineering. your agent reviews every run, writes down each lesson and adds it to its notes. it never rewrites them, so nothing it learns is ever lost
same weights, zero fine-tuning. 38.6% → 64.2%, level with IBM's top production agent on Opus 5.5. fine-tuning for that jump costs $5,000-20,000. here it cost $0
it comes down to 4 agents:
→ worker: does the task and saves the full log
→ critic: reads the log and marks every mistake and every win
→ teacher: turns each one into a one-line lesson
→ librarian: files the lesson into CLAUDE.md. merges duplicates, never deletes a line
that's why asking your agent to "clean up" your CLAUDE.md is a mistake. every rewrite throws away lessons it already paid for
paste this into Claude Code:
"Run 4 agents on my last 10 sessions in ~/.claude/projects. Critic: mark every mistake and every win. Teacher: turn each one into a one-line lesson, what happened and what to do next time. Librarian: add only new lessons to my CLAUDE.md under ## Lessons, merge exact duplicates, never rewrite or delete an existing line. Worker: rerun 2 of the failed tasks with the new lessons and save the full log. Show me the diff and both results before saving"
run it after every big session. your agent gets smarter every time and it costs $0
full playbook below ↓
5
6
284
My projects. Part 7 of 7: I’d try to fill one class first.
I nearly finished an AI coding school. I still didn’t know where I’d find students, and that’s where it stalled.
If I were starting it now, I’d try to sell one live coding lesson. Pick a topic I could teach in an evening, find people already asking about it, and see if I could fill a small class.
I’d also see if I wanted to run a second one. My transcription app made about four sales. It felt good to see the money come in, and I still didn’t want to keep running it 😅
For now I’m sticking with Nourmio. The app is out, and I’m still figuring out how people find it and what makes them come back.
Where did you find the first person who paid for something you built?
1
3
137
I pointed Nourmio at a slice of chicken pie. It called it a kouign amann. Apparently I had dessert 😅
The photo mostly showed the crust. I knew what was inside, so I wanted a quick way to tell the app and get a fresh estimate.
Now you can type “chicken and mushroom pie” and tap Update estimate. Nourmio gives you a new calorie and sugar estimate for the same portion before you save.
App Store: apps.apple.com/us/app/nourmi…
2
2
151
My projects. Part 6 of 7: a crowded market is fine with me.
Nourmio is a calorie tracker. There are loads of those already. I know 😅
After the projects I’ve talked about in this series, I’m much less interested in finding an idea nobody has touched.
If people already pay to solve a problem, I have something to work with. I can look at what they use, what annoys them, and what might make them try something else.
Tally started building a form builder in 2020, when that was already a crowded category. In June 2025, the team reported $3M ARR. It took them years to get there.
I don’t need Nourmio to own the whole market. I want to build an app people keep using and a business I actually want to run.
Competitors don’t hand you customers, though. I still have to find a reason for someone to choose us and a workable way to reach them. That’s the work I’m focusing on now.
Part 7: what I’d check before spending months on another product.
blog.tally.so/from-2-to-3m-a…
1
4
113
My projects. Part 5 of 7: the app I neglected picked up around 300 installs.
A while back, some friends and I built a mobile app and put it on the App Store. Then I kind of left it alone.
While I wasn't giving it much attention, it picked up around 300 organic installs.
That app is Nourmio. It started as a sugar tracker, then we added calorie tracking too. You can take a photo of a meal or describe it by voice, check the estimates and save it to your diary.
After spending so much time on projects I ended up closing, I had this one sitting there with people already finding it. Figured it deserved more of my attention 😅
Now Nourmio is my main focus. We're improving the app, and we've got a website with SEO articles up too. I'm working on getting more people to use it and figuring out which channels are worth putting time into.
I still need to understand what makes people come back after that first download. Getting installed is one thing. Becoming an app someone actually keeps using is the work ahead.
For now, I want to keep going with this one. Work on the product and how people find it, rather than jump straight into building something else.
Part 6: why I'm happy to build in a market that already has plenty of competitors.
6
7
235
65 for GPT-6.2 Astra is the prediction in the image. On the actual Artificial Analysis leaderboard, Gemini 4 Carbon is at 60 and GPT-6 Astra (max) is at 57. If 6.2 does hit 65, I’d want to see how much time and money it takes to get there.
artificialanalysis.ai/leader…
1
2
300
A hurricane ripping up nearly 4,000 blocks in a fan-made Rust rewrite of Minecraft is a pretty fun way to stress-test a game. I’d love to see the frame-time graph right as the buildings start coming apart.
Minecraft was rewritten from scratch in Rust.
So I made a hurricane that rips up nearly 4,000 real blocks, trees and buildings included
Rust doesn’t care how many blocks you throw at it
185
For bug fixes, I’d add this to the Claude Code prompt. “Write a test that reproduces the bug. Run it before and after the fix.” Then when it says “fixed,” you can look at the test and see what it actually checked.
This is f*cking insane. The engineer who built Claude Code just dropped a 28-minute video on how to write prompts that actually work.
I've seen $600 courses that don't cover what he shows in the first 5 minutes.
Watch it, then read the guide below on the Claude features 99% of users never find.
4
2
94
I checked Polymarket after seeing that 20% figure. The weekly market was at 37% for Nov 2–8 and 33% for Nov 9–15 when I looked on Nov 4. About $86k traded across that market.
The detail that’s easy to miss: the contract covers GPT Astra 6.1 or newer. Traders are guessing when the next Astra ships. The 6.2 name is still a rumor, and Fable has its own market.
polymarket.com/event/when-wi…
🚨 OpenAI is dropping more hints about its upcoming model
and Astra 6.2 could finally be around the corner.
Polymarket's odds of a launch today have jumped to 20%.
We might only have a few hours left before finding out.
I wouldn't count on Fable 5.5 dropping alongside it. A double release looks increasingly unlikely.
107
I get the appeal. You can spend half an hour tweaking how something moves, then play the same five seconds over and over just to see if it feels right. Curious what was the first thing in this game you got working and just kept playing with?
Working on my first game. Creating games is actually much more fun than playing them for some reason 🫣
43
Being able to build your own PDF tool is a real shift. For a payment app, here's the test I'd run: the bank took the money, your request timed out, and the customer tapped Pay again. What happens next?
1
26
I spent almost 7 months building an AI marketer. Then I realized one bad day at Yandex Direct could screw up the entire product 😅
My projects. Part 4 of 7.
The idea was pretty simple.
You tell the AI about your business. What you're selling, who your customers are, how much they usually spend.
It asks a few questions, studies your audience and competitors, figures out what kind of ads might work.
Then it builds a landing page, writes the ads, launches the campaign and keeps track of everything.
Basically, an entire marketing team in one AI tool.
And the crazy part? I was actually pretty close to finishing it.
But there was one problem.
Yandex Direct.
Some days it worked fine for me. Other days, something would go wrong.
And my entire product depended on it.
Imagine paying for an AI marketer, getting a beautiful landing page, perfectly written ads... and then nothing launches because the ad platform isn't cooperating.
What am I supposed to tell the customer? "Sorry, try again tomorrow"? 😅
That's when it hit me.
I'd spent almost 7 months figuring out how to automate everything, but hadn't properly answered one basic question:
What happens when the platform I depend on doesn't work?
So yeah, I shut the project down.
Not exactly the ending I was hoping for))
If I ever try something like this again, I'm doing it differently.
First, run the entire process manually for one real business. See what breaks, what gets stuck and what actually works.
THEN automate it.
Seven months is a pretty expensive way to learn that lesson.
Part 5: Nourmio. The project I actually decided to keep building.
1
90
Anthropic might be secretly testing Claude Fable 5.5, and the leaks are getting pretty interesting 👀
First, claude-fable-5-5 randomly showed up in the LiteLLM model catalog on October 13.
Then it disappeared the next day.
And on November 1, people spotted something even weirder: Mythos 5.5 hiding in Anthropic's own website code.
Well, that's one way to tease a new model without announcing anything 😁
Meanwhile, some Claude users believe they're already being routed to a stronger unreleased model.
There are also rumors that the next Fable could be much better at understanding what you actually want, without needing a 500-line prompt explaining every little detail.
Now THAT would be interesting.
Imagine telling Claude to build something and having it actually understand the requirements instead of confidently inventing half of them))
But here's the catch.
No official announcement. No confirmed benchmarks. No release date. And nobody has proven those impressive demos actually came from Fable 5.5.
Reddit is already debating whether better AI should guess what you mean or just be smart enough to ask when something isn't clear.
Personally, I'd love to throw it at a massive PHP backend project.
Give it a complicated task, let it work for a few hours, and see if it can actually finish without losing context or making a mess.
Because at this point, I don't need another model that can generate a beautiful demo in 30 seconds.
I need one that can survive a real codebase without giving me a headache 😅
Between GPT-6.2 Astra and Fable 5.5, things could get pretty damn interesting.
Now we just need someone to actually release one of them.
267













