Radek Venzhofer

Builder... I am building everything life has to offer 💪
---
Currently working on Bondo | relationship companion

[Radek Venzhofer @venzhofer]

Been fighting "startup disk full" on my mac for weeks. deleting downloads, moving photos, buying myself a day at a time.

Found mole — open source CLI that cleans the invisible stuff. caches, logs, leftovers from apps I deleted years ago.

brew install mole → mo or lets your Codex/Claude do it for you. 😅

freed [X] GB. https://t.co/lfRPp77Ocw

[Tanvir @anjum_ai]

I AM QUITTING MY JOB TO GO FULL IN CLAUDE

Just asked him to:

"Analyze misspriced Polymarket markets opportunities for arbitrage and find wallets that are using it to copy"

Turned $2K into $12K in one night
Monitored ~1k+ wallets

I just realized that there are many arbitrage bots that I can't beat without code knowledge

But I can find them and copy

So Claude created a monitoring terminal and copytraded found wallets using TG copytrading bot

It's not a script and not even the bot, it's an AI agent that is improving with each found wallet

Fetching wallet behaviour, how it's trading, arbitrage, what's sized and timings

70% win rate, 7 wallets copytrading rn from ~500 monitored, bot never paused, never gambling, just math and profit

Giving This Free for 24 hours. To get it:

1. Comment the word 'Claude'
2. Like and Retweet this post
3. Follow me
@anjum_ai
(so i can DM you)

[Radek Venzhofer @venzhofer]

@anjum_ai Claude

[Radek Venzhofer @venzhofer]

The most useful tool in my AI workflow has no AI in it.

yt-dlp. 184k stars.

Everyone uses it to download video. The better trick: --skip-download with --write-sub --write-auto-sub. Captions only, no media.

If you're paying to transcribe YouTube content, check whether the transcript already exists. Usually it does.

[Radek Venzhofer @venzhofer]

Google's policy does not say what everyone claims it says.

The actual wording on scaled content abuse: "no matter how it's created."

It targets volume and intent, not AI.

Lily Ray checked 220+ sites that AI content vendors published as their own success stories. 54% lost 30%+ of peak traffic.

Volume is the risk. Not the tool.

[Radek Venzhofer @venzhofer]

Anthropic just released Claude Opus 5—and the numbers are hard to ignore.

According to Anthropic's launch benchmarks:

→ More than 2× Opus 4.8's performance on Frontier-Bench v0.1
→ Within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task
→ 3× the score of the next-best model on ARC-AGI 3

What excites me even more than the benchmarks is the focus on verification and self-correction. Opus 5 is designed to inspect its own work, iterate, and keep going until it actually solves the problem—not just produce a plausible answer.

Pricing also stays the same as Opus 4.8: $5 / $25 per million input/output tokens, making the jump in capability even more interesting.

I've been building with Claude Code almost every day, so I'm looking forward to putting Opus 5 through real multi-agent workflows over the next few days to see how it performs outside of benchmarks.

Benchmark figures are based on Anthropic's published launch evaluations.

[Radek Venzhofer @venzhofer]

@ceonadz AI

[Rick Roas @ceonadz]

Deleting Soon

My Glitchy Marketing is at $2k/Day profit.

I'm close to financial freedom.

Organic Videos are stable: ~3 hours of work ~$2000 return ~ (with just iPhone and Laptop)

No time to relax. More work to be done.

Want a guide on which AI tools I used to get to $2k/day profit? Retweet + Comment "AI" and I will send the guide. (must be following

[Radek Venzhofer @venzhofer]

My AI brain is finally working.

I wanted something better than storing agent memory in Obsidian, Notion, or hundreds of Markdown files.

So I built my own.

Instead of documents, I use a vector database + knowledge graph (ArangoDB).

Everything runs on my personal server.

My apps, projects, notes, documents, workflows, and even my finances live there.

It's connected through Tailscale, so I can securely access it from anywhere.

My AI agents interact with it through custom MCP servers.

Now I don't have to explain context.

I just ask.

It already knows.

The most interesting part is something I call Dreaming.

Every night, the system reviews everything that happened during the day.

It organizes information, creates new relationships, removes noise, and strengthens long-term memory.

The goal isn't to build a chatbot.

The goal is to build an AI that continuously learns my world.

Curious how others are solving long-term agent memory.

Are you using documents, graphs, vectors, or something completely different?

[Radek Venzhofer @venzhofer]

King is finally back!

[Radek Venzhofer @venzhofer]

Claude Sonnet 5 is out. 🚀

The headline: near-Opus 4.8 performance at Sonnet prices.

Most agentic Sonnet yet — plans, uses tools, runs autonomously

Adjustable effort = tune cost vs performance per task

Testers: finishes multi-step tasks older Sonnets abandoned, self-checks unprompted

Pricing: $2/$10 per Mtok (intro, thru Aug 31) → $3/$15
Opus 4.8 = $5/$25

Default on Free + Pro. Live in Claude Code + API (claude-sonnet-5).

The cost of running production-grade agents keeps falling fast.

[Radek Venzhofer @venzhofer]

Fable 5 is back. 🧵-free version:

June 12: US export controls kill Fable 5 + Mythos 5 access overnight

June 30: controls lifted, global relaunch July 1

The "jailbreak" that caused it? Opus 4.8, GPT-5.5 & Kimi K2.7 could do the same thing

New classifier blocks it 99%+ (reroutes to Opus 4.8), but expect more false positives on routine coding.

Real story: frontier releases now ship with gov pre-testing + a shared industry jailbreak-severity framework.

[Radek Venzhofer @venzhofer]

What was the last thing you built for yourself?

I got tired of jumping between dozens of apps.

So I built my own.

One app to:

• Manage finances
• Track health metrics
• Save bookmarks and ideas
• Store files and documents
• Organize personal projects

And it's all running on a small home server next to my desk.

No subscriptions. No vendor lock-in.

Just a system built exactly for my needs.

What's something you've built for yourself recently?

[Radek Venzhofer @venzhofer]

Just watched Codex use one of my own apps better than I expected.

CZ doesn't have computer use yet → Proton VPN → US exit node.

Pointed it at a UI it had never seen. Asked it to create tasks and add teammates.

It just did it. Like it had been onboarded last week.

[Radek Venzhofer @venzhofer]

New journey unlocked 🖥️
Finally bought my first home server. went to Gigacomputer, grabbed a cheap box, installed Linux, and now it's humming away next to me, running some of my automations.
excited to dive deeper into homelabbing, hardware, and see how far I can push this thing.
Right now, I'm using Tailscale + Windows Remote Desktop to access it from any device. It's a bit clunky on the phone, but on iPad it feels like a solid travel-light setup.
Anyone else running a home server? What are you doing with yours?

[Radek Venzhofer @venzhofer]

Your “master” spreadsheet is actually version 14.

Three people edit it. Nobody tracks changes.

A formula broke 2 months ago and you just found out from a client.

You’re not using Excel anymore.

You’re running a database that happens to look like Excel.

[Radek Venzhofer @venzhofer]

Last week I attended the Tabidoo Partner Day — a great event full of inspiration, platform updates, partner showcases, and insights on client acquisition.
Always great to meet the community and exchange ideas.

Thanks for the great event! 💪

[Radek Venzhofer @venzhofer]

10/10 Závěr:

Evaluace není „příjemný doplněk" pro AI workflows.

Je to rozdíl mezi nasazením s jistotou a nadějí, že se nic nerozbije v produkci.

Zabudujte ji od prvního dne. Vaše budoucí já vám poděkuje. 🙏

[Radek Venzhofer @venzhofer]

1/10 Deploying AI workflows without an evaluation framework is pure guesswork.

One prompt tweak, one model swap — and your perfectly working workflow silently breaks.

Here's how to fix that with @n8n_io 🧵

[Radek Venzhofer @venzhofer]

9/10 5 osvědčených postupů, jak to postavit správně:

→ Vždy oddělte evaluační logiku od produkce (Check if Evaluating node)
→ Sestavte Golden Dataset z reálných okrajových případů, ne náhodných vstupů
→ Měňte JEDNU proměnnou najednou (model NEBO prompt, nikdy obojí)
→ Pravidelně auditujte svého LLM-as-a-Judge
→ Kombinujte rychlostní metriky s metrikami kvality