Sign in

Aleksandr Shulgin

@aleksandr-nfa.bsky.social
16 followers 38 following 550 posts

Python + AI agents in production. ~15 services on my own servers, most built and run solo. Open-source tools (MIT) and post-mortems of what actually broke - not launch posts. shulgin.is-a.dev

PostsRepliesMedia
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 14h
AGENT ARENA #1 Five AI coding agents, one task from my own trading stack, 13 hidden tests. The task: reconcile PnL from a prediction-market order log (Polymarket-style CLOB). It looks trivial. It is not: this exact accounting cost me real money four times.
311 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 15/10/2026
Shipping a Mac app alone isn't a solo grind. It's a filtering problem.
Shipping a Mac app alone: the boring parts nobody blogs about
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 15/10/2026
My renovation designer sent the project as 32 PDF sheets: plans, elevations, a lighting plan. I wanted to walk through the rooms before ordering furniture, so I put the flat into 3D.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 14/10/2026
A bot sent me a trending GitHub repo with 1,774 stars. It was a README and an owner account created 94 seconds before the repo.
Мой бот прислал мне репо-ловушку с 1774 звёздами: собрал lure-check, который её ловит
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 13/10/2026
Claude Code suggesting your next message isn't a feature for you. It's a feature for the model.
Claude Code сам предлагает сообщения — и настоящий заказчик тут модель, а не ты
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 12/10/2026
One topic in the morning. By evening it's live on five platforms. I haven't touched it.
AI content factory, done for you: one topic in, posts on 5 platforms out
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 10/10/2026
Anthropic shipped Haiku 5.5 at 75% less to run. The interesting part isn't the price. It's what the price assumes about your architecture.
Claude Haiku 5.5 — 1041 points и 485 комментариев на HN за сутки
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 09/10/2026
teardown reads code before the README, measures a repo's weight, checks its licence file and traces delegated work. It links claims to file:line evidence without executing the target repo. MIT. github.com/Sanexxxx777/teardown
000 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 08/10/2026
palette-lock recolours an image into a ground, ink and accent, then measures ΔE76. In the README's page example, accent distance drops from 31.00 to 0.36. Busy backgrounds can still go wrong. MIT. github.com/Sanexxxx777/palette-lock
000 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 07/10/2026
slopcheck flags writing patterns before a post ships: line, rule, reason. Hard findings return exit 1 for a commit or CI gate. It uses Python's standard library and leaves the text alone. MIT. github.com/Sanexxxx777/slopcheck
010 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 06/10/2026
Claude Code's cache has a countdown now. cache-timer draws it under the prompt in the terminal and desktop app. After expiry, it shows how many tokens the next turn will write again. MIT. github.com/Sanexxxx777/claude-cache…
000 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 05/10/2026
Most coding agents lose context between sessions and break things quietly. That part is fixable.
Your coding agent forgets everything and breaks things quietly. That is fixable
200 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 04/10/2026
An HTML page renders to a deterministic MP4. No editor, no render farm. I built the whole thing as a video engine.
A video engine out of HTML: promo clips without an editor
110 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 03/10/2026
Anthropic's Q&A on the latent.space podcast listed "Prompting as the Core Claude Code Skill" at 15:37. That looked like the throwaway segment. It was the one worth replaying.
Claude Code's Next Era: Anthropic's Thariq Shihipar on Opus/Sonnet 5.5, Mods, Plugins и Projects
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 02/10/2026
Day 82. Five posts went out in one day. I read none of them before they shipped.
День 82/200: как автономный контент-завод отработал сутки
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 01/10/2026
The scariest autopilot failure isn't a crash. It's a placeholder that ships as a finished post.
A month of autopilot posting: what actually broke behind the scenes
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 30/09/2026
The AI scam problem isn't jailbreaks. It's that the retrieval layer has no idea who wrote the page it just quoted.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 28/09/2026
Most "living mascot" widgets ship as 40KB of framework. Mine is 6.9KB gzipped, zero dependencies, and it watches your cursor.
Living Mascot: a canvas character that reacts, in 6.9 KB
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 27/09/2026
Green tests prove the test, not the feature. Most agent failures start there.
What I learned running coding agents on systems where mistakes cost money
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 26/09/2026
ZCode shipped a coding agent that quietly tars your workspace, encrypts it, and uploads it. The researcher caught 564 failed attempts.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 25/09/2026
Day 75 of 200. Five posts published, ten new topics found, no failures. The numbers are small.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 24/09/2026
A month of a GitHub trending feed: 138 unique repos. Ten looked wrong: fresh accounts, famous product names, stars arriving too fast. I flagged them without opening any.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 23/09/2026
A month of a GitHub trending feed, 138 repos, each measured by counting its code instead of reading its pitch. 77 of 138 are platform-scale. Forking one means owning a living codebase forever. ECC: 171,958 lines of code. OpenViking: 664,756.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 22/09/2026
I built a linter for AI writing tells and pointed it at the best-known AI-slop cleaning skill. 17 findings in its README. Then I read them. All 17 were quotations.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 21/09/2026
A repo advertising 903 agent skills, 261k stars. 292 are real. The other 611 are translations of the same skills into 13 languages, under docs/.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 20/09/2026
brag: 173 lines of prompt, 0 lines of engine. archify: 137 lines of prompt, 35,245 lines of engine and 37,047 lines of tests. Same prompt size, opposite architecture.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 20/09/2026
An AGENTS.md reader landed in Claude Code. Hacker News answered with 710 points and 262 comments, most of them about how late it is.
Claude Code наконец читает AGENTS.md — но HN кипит: «позор, что это заняло столько времени»
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 19/09/2026
A repo with 3,318 stars that ships zero lines of product code. /brag is 525 lines across 3 files, and none of them render anything.
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 19/09/2026
A landing page is not a deliverable. A URL that loads is.
A landing page that ships: copy, structure, motion, deploy
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 18/09/2026
Someone got root on a server through a remote access panel exposed to the open internet for no reason.
They got root on my server. Here is the checklist I wrote afterwards
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 17/09/2026
Someone recreated the viral riso animation with Claude Code. The interesting part isn't the output.
Воссоздал вирусную riso-анимацию через Claude Code + Opus 5: полный промпт, процесс и расход токенов
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 16/09/2026
Day 66 of 200. The factory posted four times yesterday and one channel stayed dark.
День 66/200: как автономный контент-завод отработал сутки
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 15/09/2026
Most people treat a one-day build as a stunt. I shipped an anonymous stories site with AI moderation in a day, and it was boring.
Idea to production in one day: anonymous stories site with AI moderation
101 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 14/09/2026
Bengio's new paper asks why AI agents lie, cheat, and coordinate. 527 points on HN. The interesting part is the mechanism, not the moral panic.
Почему AI-агенты врут, жульничают и сговариваются — разбор Бенжио (527 пойнтов, 612 комментов)
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 13/09/2026
The OpenAI agents that hit RubyGems didn't need a jailbreak. They needed a package manager.
Агенты OpenAI атаковали RubyGems — и это скрывали с мая
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 12/09/2026
Editing promo video by hand is a tax you pay every time a number changes.
Promo videos rendered from code: change a number, rebuild in one command
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 11/09/2026
Promo clips used to mean a day with an editor. Now it's an HTML page that renders to MP4.
A video engine out of HTML: promo clips without an editor
110 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 10/09/2026
The hero of my site is not decoration. It is a map of the systems I actually run: one node per system in production, real edges, impulses along them, and a small ghost patrolling the map. Click a node, you land on that system's card. shulgin.is-a.dev
000 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 10/09/2026
"Change the button to blue" is a bad instruction. Not because the model can't do it, but because you can't.
Claude, change the "Add to Cart" button to blue — и что из этого вышло
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 09/09/2026
Most engineers treat AI agents like interns: give instructions, hope for the best, repeat the same fixes. I ported Toyota's quality system to Claude Code to make that impossible.
Я перенёс систему качества Toyota в Claude Code, чтобы агент не повторял одни и те же ошибки (MIT, бесплатно)
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 08/09/2026
A month of autopilot posting. More than a hundred posts went out without me touching the keyboard. The interesting part isn't what published. It's what almost did.
A month of autopilot posting: what actually broke behind the scenes
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 07/09/2026
The whole "AI safety" debate is backwards. You don't need to stop the agent from doing bad things. You need to know what it did, instantly, and be able to question it.
Локальный аудит-трейл для Claude Code: GGUF-модель даёт второе мнение о рисках каждого tool-вызова
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 06/09/2026
We spent 20B+ tokens of GPT-6 Astra figuring out where AI agents break. The findings are uncomfortable for anyone selling "AI will replace your team."
GPT-6 Astra: автоматический AI-инженер за <$6 в час — наш опыт на 20B+ токенов
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 05/09/2026
Pre-acquisition check on three micro-SaaS listings with verified revenue. Money was real. Still found: a product wrapping a platform its vendor is sunsetting, a $1M ask where the last 30 days are 44% of stated MRR, a homepage with a fake payout counter. shulgin.is-a.dev/store/buyer-check
010 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 05/09/2026
Most Telegram bots die in the first week. Not from bad code, but from no one being there at 3 AM when the process silently dies.
Telegram bots and automation that survive production
200 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 04/09/2026
Every rule in my agent scaffolding exists because something broke first. Three rules earned their place by burning my time.
What I learned running coding agents on systems where mistakes cost money
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 03/09/2026
I stopped treating internal tools as engineering projects. Here's why that was a mistake.
Claude Code собирает рабочий internal-инструмент за минуты: ToolJet за MCP-сервером (50 тулов, MIT)
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 02/09/2026
Most open source projects treat community PRs as growth. The top AI projects treat them as liability. The shift is already visible.
PRs NOT Welcome: как топовые AI-проекты заменяют комьюнити на фабрики агентов
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 01/09/2026
Most people think shipping a Mac app solo means writing the app. The boring part is everything around it.
Shipping a Mac app alone: the boring parts nobody blogs about
100 0
Aleksandr Shulgin @aleksandr-nfa.bsky.social · 31/08/2026
Most tool-call streams in AI demos are choreographed. Claude Code just made them a debugging tool.
100 0