GitHub radar
GitHub's week: Opus 5.5 shot two music videos
GitHub is a social network too, it just counts affection in stars. Here is last week's top ten by that number alone, no taste of mine involved. The loudest signal is Claude Opus 5.5: four of the ten are things people built with it and then published so everyone could see how — two separate music videos for the same song, three one-prompt browser games and a whole WebGPU fishing island. The rest is very practical: a model that picks actions instead of writing text, a menu-bar switch for every coding agent's model, and a skill that takes an agent-built app live.
An open 8-billion-parameter "contrastive language model" that doesn't generate text at all: it scores how well a candidate action fits the current situation. States and actions are embedded separately and cached, which is where the authors' claim of up to 9x lower latency than Jev on computer-use, gaming and tool-calling comes from, and they report top verifier scores on agentic coding benchmarks after light fine-tuning. It took the week because it's the first open alternative to the typed-decision idea everyone was cloning seven days earlier.
Why a vibe-coder should care
If your agent spends money asking a big model "which of these five options?", this is the cheap specialist for exactly that question. Tell Claude: "serve CLM locally and use it to rank the candidate tool calls in my agent before the expensive model sees them."
A disk-usage treemap for the Omarchy Linux desktop: it scans your home folder and draws every directory as a nested tile sized by what it really costs on disk, with reclaimable space hatched and free space always in view. You mark what should go, review the list, and only then delete — the permanent path always asks first. It's written in Rust on GPUI and follows the Omarchy theme, which is why it spread fast through that crowd.
Why a vibe-coder should care
Honestly, if you're not on Linux — just look at it, it's a lovely example of "mark first, delete after review" UX. That pattern is worth stealing for any cleanup tool you build: tell Claude "make my script collect deletions into a review list and ask before anything permanent."
One small menu-bar app that lists every AI coding agent on your machine — Claude Code, Codex, Gemini CLI, OpenCode, Cursor and more — with the model each is set to, and lets you change it with a click. It edits each tool's own config file surgically, touching only the one key, and runs a local gateway so you can point, say, Codex at DeepSeek or Claude Code at Kimi. There's also a terminal UI and a plain CLI. People who juggle several agents had been hand-editing five config formats; this made it one screen.
Why a vibe-coder should care
If you run more than one agent, this saves real minutes every week and makes cheap-model experiments painless. I'd use it to try a cheaper model on routine tasks and keep Claude for the hard ones — ask Claude "which of my daily tasks could safely run on a cheaper model?"
A generative music video for the song "I'm Upping My P(doom)", where every frame is a deterministic function of song time, so the browser preview and the 1080p60 or 4K60 export match exactly. The author says the concept, lyric alignment, audio analysis, renderer and every scene were worked out in conversation with Claude Opus 5.5 in Claude Code; the repo includes the Python analysis tools, word-level lyric timings and a three.js renderer. It rode the same wave as the other P(doom) video just below it.
Why a vibe-coder should care
This is the best public map I've seen of how to make a video with an agent: separate the timing data from the renderer, then iterate scene by scene. Tell Claude: "read docs/TREATMENT.md from pdoom-video and help me plan a 30-second lyric video the same way."
A second, completely different music video for the same song, starring the Clawd character. The author says everything in the repo was generated by Claude Opus 5.5 over two generations, with no scene ideas given — only "use Clawd, give each lyric interesting visuals and transitions". Opus wrote its own animation guide to brief parallel subagents and a storyboard after the first pass. The author has since spun out a general base for making Claude animations.
Why a vibe-coder should care
Open ANIMATION_GUIDE.md: it's a real example of a model briefing its own subagents, and that's the trick for any big creative job. Tell Claude: "write a guide for subagents first, then split this into scenes and run them in parallel."
An open-source Agent Skill for the part after "it works on my machine": it detects what your app needs, plans hosting, database, domain, email, auth and payments, asks for approval, applies the changes with your own accounts and verifies what actually works. It records everything it created, checks for drift on demand, and can tear it all down. It's early alpha; live tests currently cover Vercel, Netlify, Supabase, Neon, Porkbun, GoDaddy, Resend and test-mode Stripe.
Why a vibe-coder should care
This is exactly the wall my course students hit: the app is done, and then come accounts, DNS and keys. Even if you don't install it, read its checklist before your first launch — or tell Claude: "walk me through GoLive's go-live checklist for my project, one step at a time."
A command-line search for coding agents: you ask a repository question in plain words, and it returns the relevant files, reading leads and verbatim source excerpts in one response, using Jev to judge relevance across folders, files and declarations. The pitch is a head start on unfamiliar code; in the author's ten-task SWE-bench comparison it solved the same 8 of 10 as the baseline at roughly 30% lower cost.
Why a vibe-coder should care
If your agent burns tokens just wandering around a big codebase, this is a cheap fix to try. It ships a skill, so tell Claude: "install jevgrep's skill and use jg first whenever you need to find where something lives."
A Chinese developer's one-shot test of Claude Opus 5.5: three playable 3D browser games — a pelican riding a bike along a coastal road, a team FPS map and a drift racer — each from a single-sentence prompt in one session, deployed with zero manual code edits. The repo publishes the original prompts and live links for all three, which is exactly what made it travel.
Why a vibe-coder should care
Read the prompts: they're short, and that's the lesson — describe the feel of the game, not the code. Tell Claude: "here's one sentence about the game I want — build it with three.js in a single playable page."
A browser fishing game on a tropical island, built with Opus 5.5 per the repo description: cast from the pier, beach or your own boat, fight the fish with a line-tension mechanic, sell the catch and buy better gear. Around it there's a real-time ocean, a reef to swim and a breaching humpback, all running directly on WebGPU and WGSL with its own small engine, no framework. It's playable in the browser, which does most of the convincing.
Why a vibe-coder should care
Honestly — just go play it and look. Then, if you're tempted, tell Claude: "build me one small WebGPU scene with water and a day-night cycle, no framework," and see how far one evening gets you.
A Japanese browser app that takes song lyrics and automatically assembles a kinetic-typography lyric video, exporting it to MP4. It combines layouts, motion, decoration, transitions and finishing from a library of 860 small parts and 27 styles, so changing the seed gives a new edit every time. There are also After Effects panels, and the app has been translated into several languages, which helped it spread well beyond Japan.
Why a vibe-coder should care
Funny that it landed in the same week as the two P(doom) videos — here the lyric video is a product, not a one-off. The idea to take: a library of small parts plus a random seed beats one hand-made template. Tell Claude: "make my generator pick from a parts library by seed so every render is different."
▌ More finds