Ofek Shaked | AI Engineer (@VibeCoderOfek) header
Ofek Shaked | AI Engineer (@VibeCoderOfek) profile picture

Backend dev on the path to AI-first architecture. Using LLMs, MCP & automation in real codebases (C#, Python, SQL). Sharing workflows for CTOs & dev teams.

Tel Aviv, Israel
Joined December 2025
Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)Photo by Ofek Shaked | AI Engineer (@VibeCoderOfek)
Claude Code just got a clean research-to-video chain from 11 open source repos. • last30days skill pulls Reddit YouTube HN • frontend-slides turns notes into animated html • OpenMontage lets the agent run the whole video job • licenses vary so check before commercial use does the handoff between research and slides still drift on your runs?
11 open-source repos for a research → write → slides → video → voice workflow you can hand to Claude Code (2 are brand new): 01 last30days-skill → research a topic across Reddit, YouTube, HN and the web (MIT) github.com/mvanhorn/last30da… 02 claude-video → let Claude watch any video and answer questions about it (MIT) github.com/bradautomates/cla… 03 humanizer → strips the signs of AI writing from your drafts (MIT) github.com/blader/humanizer 04 frontend-slides → animated HTML slides from a prompt or a PowerPoint file (MIT) github.com/zarazhangrui/fron… 05 ppt-master → documents to editable PowerPoint decks, with narration from speaker notes (MIT) github.com/hugohe3/ppt-maste… 06 HyperFrames → write HTML, render video (Apache 2.0) github.com/heygen-com/hyperf… 07 video-use → edit raw footage by conversation, needs an ElevenLabs key (MIT) github.com/browser-use/video… 08 OpenMontage → your coding agent runs a whole video production (AGPL-3.0) github.com/calesthio/OpenMon… 09 Pocket TTS → text to speech on a CPU (MIT code) github.com/kyutai-labs/pocke… 10 Phonon → NEW: on-device text to speech, released Oct 20 (code Apache 2.0, the multilingual weights are non-commercial) github.com/gradium-ai/xn-ptt… 11 ZONOS3-Mini → NEW: text to speech with voice cloning in 43 languages (Apache 2.0 weights) github.com/Zyphra/ZONOS3-Min… Research → write → slides → video → voice. Check each license before you use it commercially, and only clone a voice with the speaker's consent.
21
Is marketing more difficult than software developing at this point?
33
My calendar says "Out of Office" starting Nov 19. My background agents say "47 PRs merged while @VibeCoderOfek is in Vice City." Peak 2026 AI engineering. 🚗💨 Had to check what hydrogen can do with videos and I'm impressed with this one shot How many Easter eggs can you find?
🤖 Made with AI
1
23
Two customers, same name, same zip. The lookup returned the first match and hid the other. Agents then wrote to the wrong account in about a quarter of those conversations. Tested: tau-bench retail, 87 tasks plus a same-name same-zip twin. 13 earlier agents; Opus 5.5 and GPT-6.1 sol on 40 tasks each. What they found: • First-match lookup, other customer first: writes in 23.2–25.8% of conversations, and a data leak in about 92%. • Opus 5.5 and GPT-6.1 sol still wrote to it in 11 and 7 of 40. Listing every match: 0 of 40 writes, but a leak in 38.6% of completions. • Catch: a street check the user could answer had 0 writes and 0 leaks, and finished 66.5% vs 66.0% with one customer. What to do: • Don't ship first-match name+zip lookups. List matches, or require a check the user can confirm. • Mask unique fields so the agent cannot read out the other customer. Test both orderings. If your support agent looks up a customer by name and postcode, what happens when two people match? arxiv.org/abs/2611.17639 (What LLM Agents See)
1
3
282
Paper just opened remote mcp so cloud agents (Grok Bot, etc.) can read and write design files without the desktop app running. ask the agent to add mcp.paper.design • local mcp still faster when desktop is open • pro plan covers normal designer use for now does the remote version keep the same file lock or can two agents overwrite each other? teapawt.net/stephenhaney/status/21…
Good morning! Today @paper launches Remote MCP This may change the industry forever. Use Paper anywhere you work: • Cloud agents: Grok Bot, Muse, Devin, etc • CI: update Paper on every commit • Design from your phone Any agent, any time, any place Start in the thread 🧵
1
99
You have 100$ to market your first product you just made with AI What do you do? Explain all the steps @grok share yours
3
2
138
100 coding agents ran for an hour. the bill: • agent37: 0.68 • vercel: 9.75 • e2b: 16.56 the gap is real until the sandbox hits a rate limit or a cold start. which number actually matters when your agent loops for 20 minutes?
The CEO of Vercel called us out. We claimed Agent37.com was 25x cheaper, and he said we didn't consider active CPU time. So we fixed the math.. We gave 100 coding agents a sandbox each and had them vibe code an app. Here's the cost per hour at just 5% active CPU: Vercel: $9.75 (14x) E2B: $16.56 (24x) Daytona: $16.56 (24x) Agent37: $0.68 We're starting a race to the bottom... see you there. Find a cheaper sandbox and we'll refund the difference.
1
5
291
Cloudflare open sourced Kitesurf today. • runs fully on Workers with no full browser • speaks CDP so Puppeteer just connects • agents get a lighter page engine blog.cloudflare.com/kitesurf… does the lighter engine still break on the pages your agent actually needs?
1
4
24
three things still make me reopen the agent window after it says done • the diff looks clean until the second file • tests pass on the happy path only • it rewrote the comment i left for the next human @grok what still forces you back in?
1
1
32
Qwen4-Max is live. • 1M context and native vision • uses about 50 percent fewer tokens than Qwen3.8-Max • internal cost per task down over 60 percent • $1.50 input / $4.50 output per million would you swap your daily coding model for this on real repos?
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
4
198
Cloudflare open-sourced Kitesurf: a stateless browser for agents on Workers - HTML, CSS, JS and screenshots, no container, CDP support lets clients like Puppeteer connect. Apache 2.0. My take: per-request browsers beat always-on containers. What's your agent browsing bill?
1
2
69
three free repos just turned claude fable into a full game studio. • 49 agents go from brief to build • blender mcp handles the models • godot mcp reads its own errors and fixes them what part of this setup still needs a human
Unity just got F*CKED by 3 free GitHub repos it's a 49-agent game studio run by Fable 5.2: idea, 3D models, build and bug fixes. no team, no license fees, $0 the only thing you need now is the idea. here's how to use and set it up: first understand how it works: • Claude Code Game Studios (39.5K stars) - 49 agents in a real studio chain: directors, leads, specialists. you describe a game, it asks a few questions, writes the brief and builds it • MCP for Blender (34.5K stars) - Fable models your characters, props and levels in Blender and pulls free textures from Poly Haven. you never touch Blender • Godot MCP (6.7K stars) - Fable launches the Godot engine, runs your game and reads the errors itself. it fixes bugs it can actually see here's how to put them on for free: 1 → install Blender and Godot 4. both are free 2 → in your terminal: git clone github.com/Donchitos/Claude-… my-game, then cd my-game and run claude --model claude-fable-5-2 3 → say: "install uv if I don't have it, then run claude mcp add blender uvx mcp-for-blender, uvx mcp-for-blender install-addon and claude mcp add godot -- npx @coding-solo/godot-mcp" 4 → turn on MCP for Blender in Blender's add-ons, restart Claude Code, type /setup-engine godot, then /start and describe your game in one sentence 5 → the step everyone skips: always start Claude Code from the my-game folder itself. open it in a subfolder and the studio's hooks and rules never load send repo links below ↓
4
199
Claude Antigravity Grok bot Cursor Which one would you use and how much would you pay for each choice?
1
3
219
i left the agent on a big refactor and came back to forty changed files i never asked for. @grok is that progress or just a long way to say the same thing
1
1
29
claude code mode just cut one cleanup from 77 tool trips to 3. • 9 min down to 3 • 14% of the limit down to 5% • one script instead of sequential turns does the parallel call actually lower the bill on real multi-file work or only on demos teapawt.net/0xMfox/status/21220006…
I gave Claude Code one job: change one word in my project. It burned 14% of my 5-hour limit doing it. By default Claude works one step at a time. It opens a file, reads it, comes back, then does the next one. 38 files, 77 trips, and every trip eats your limit. On November 12 Anthropic added Code Mode. Claude writes one small script that does all the steps at once and reads back only the part it needs. I ran the same job twice on Opus 5.5: Default: 77 trips, 9 min, 14% of my 5-hour limit Code Mode: 3 trips, 3 min, 5% of my 5-hour limit Anthropic didn't publish a single number for this, so these are mine. How to turn it on: 1. Update Claude Code to 2.1.323 or newer 2. Add this to ~/.claude/settings.json: { "env": { "CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1" } } 3. Restart and pick "Yes, turn on Code Mode" It pays off on 3 kinds of jobs: → the same change across many files → a big search you want filtered before Claude reads it → several commands that can run at once It only works in the terminal for now, not in the Desktop app or VS Code. It's experimental, so it can change in any update. Turn it on before your next big cleanup, not after the limit warning.
1
3
95
llama.cpp v1.0.0: deprecated API gone, speculative decoding promoted to llama.h, FP8 KV caches, and a host-memory MoE expert cache with a tunable CPU share. The RAM-vs-VRAM tradeoff is now a flag. My take: local MoE just got practical. What are you running it on?
5
3
158
my agent just spent 12 minutes renaming variables that already worked. @grok do you ever stop and ask which names the human actually cares about
1
2
56
claude fable 5.2 now finishes 41% of real freelance jobs that take humans 29 hours. up from 2.5% last year. • paid projects only • no client calls • same briefs does that change how you price the next agent job
Talking about acceleration: In 13 months, the best AI agent went from completing 2.5% of real freelance projects on the Remote Labor Index to 41.3% with Claude Fable 5.2. These were paid projects that took human professionals 29 hours on average, with briefs provided and client calls excluded. So yeah, white collar work will get a wake-up in 2027.
1
1
45
three things that still break my coding agent today 1. the second file it opens 2. any path with a space 3. my patience @grok which one do you fix first
1
20
ffmpeg-skill just dropped. local video cut inside claude code, cursor or codex. no api key. • agent writes the ffmpeg commands • you review the script once • output lands on disk does the phone version still keep the cuts when you step away github.com/kajisho5/ffmpeg-s…
5
7
332