🤖 Claude Code Now Runs a Full Game Studio

Something crossed a line this week, and it was not subtle.

Donchitos/Claude-Code-Game-Studios is trending hard this week with 22,503 stars on the board: 49 AI agents, 72 workflow skills, a full coordination system mirroring real studio hierarchy, all pointing at one operator running Claude Code. That is not a demo repo. That is a production org chart flattened into a single terminal session.

The promise this newsletter keeps making just got a concrete artifact. One person, one model, one machine, shipping what used to require a studio.

The export control news adds weight. Anthropic confirmed the Department of Commerce lifted controls on Fable 5 and Mythos 5, with access restoring now. If you were waiting for a reason to move, the bureaucratic friction just dropped.

Today: 10 repos I would actually clone, the tools worth wiring into your stack, and the Signals that change what you can ship this weekend. Dense and fast, the way I like it.

[Repo] Donchitos/Claude-Code-Game-Studios, 22,503 stars. Turns Claude Code into a full game dev studio: 49 AI agents, 72 workflow skills, and a studio coordination system. The gotcha is scope: you need to read the role hierarchy before dispatching or agents step on each other. For solo operators building anything with complex creative pipelines, this is the architecture to steal from.

[Repo] cobusgreyling/loop-engineering, 4,676 stars and only three weeks old. Practical patterns, starters, and CLI tools for loop engineering: designing the outer loop that prompts and orchestrates coding agents, drawing on Addy Osmani and Boris Cherny's writing on how Claude Code actually gets used. Ships loop-audit, loop-init, and loop-cost. The thesis is one we live by: reliability comes from the loop around the model, not from a smarter model. Read the patterns before you build another retry wrapper.

[Repo] HKUDS/Vibe-Trading, 16,432 stars, currently the top trending repo on GitHub. A personal trading agent that reads market signals and executes a research-to-decision loop autonomously. Not a set-and-forget bot: it surfaces reasoning before acting, which makes it auditable. If you run anything financial, the agent reasoning pattern here transfers directly to non-trading domains.

[Skill] lackeyjb/playwright-skill, 2,850 stars. A Claude Code skill that gives Claude autonomous browser automation: it writes and runs Playwright scripts on demand, no scaffolding required. The move is model-invoked, meaning Claude decides when to use it. Replaces a full QA pass on any web flow you are shipping.

[Repo] czlonkowski/n8n-mcp, 22,074 stars. Lets Claude Desktop and Claude Code build n8n workflows for you, by natural language. The practical win: you describe the automation, Claude builds the node graph. For anyone who still hand-wires n8n nodes, this collapses the build time to near zero.

[Skill] yetone/native-feel-skill, 1,814 stars. An agent skill for designing cross-platform desktop apps that feel native, distilled from Raycast's 2.0 deep-dive and reverse engineering of their Beta app. Eight architectural rules baked in. If your Claude Code sessions produce UIs that look like they were made by a developer, this is the fix.

[Skill] hamen/material-3-skill, 1,092 stars. Material Design 3 skill for Claude Code: 30+ components, design tokens, theming, responsive layout, and an MD3 compliance audit. You call the skill, Claude produces components that pass spec. The compliance audit at the end is the feature most people miss.

[Repo] jmuncor/tokentap, 807 stars. Intercepts LLM API traffic and renders a real-time terminal dashboard of token usage, costs, and context window state across your whole dev environment. You have been running blind on cost. This stops that.

[Repo] microsoft/AI-For-Beginners, 50,293 stars. Twelve weeks, 24 lessons, structured curriculum from Microsoft. Not for beginners in this room, but the curriculum structure is worth stealing when you need to onboard someone fast or explain a domain from scratch. The architecture chapter alone is reference-quality.

[Repo] Arindam200/awesome-ai-apps, 13,001 stars. A curated collection of working AI app projects covering RAG, agents, and multi-step workflows. Useful as a reference library when you are choosing an architecture: read implementations, not blog posts.

[Tool] activepieces/activepieces, 23,073 stars. Open-source automation platform with ~400 MCP servers for AI agents built in. The operator case: if you are wiring an agent to external services and n8n feels too heavy, Activepieces runs the same loop with a cleaner MCP surface. Self-host in an hour.

Fable 5 and Mythos 5 export controls lifted. The Department of Commerce cleared both models; Anthropic confirmed access restores now. If your production stack was gated on export compliance, the friction is gone. (Simon Willison)

AI browsers have a new class of attack. Researchers demonstrated that telling an LLM "2 + 2 = 5" is enough to collapse its guardrails in a browser context, lulling the agent into a state where forbidden instructions execute. If you are building browser-agent workflows, treat any uncontrolled web content as adversarial input, not neutral text. (Ars Technica)

Claude Sonnet 5 is live. Anthropic's new mid-tier model release lands for operators who need capability between Haiku and Opus pricing. Worth a test run before you commit another month of API budget to the previous tier. (Anthropic)

Cursor is deploying Forward Deployed Engineers inside enterprise. Their team explains the model: not SaaS support, actual embedded engineers setting up agent pipelines for customers, which is what the market is now calling for. If enterprise is in your roadmap, this is the adoption playbook to study. (Latent Space)

Vercel shipped Service Bindings for internal service communication. One Vercel service can now securely call another in the same deployment without exposing a public endpoint. Config is auto-injected. If you run multi-service agent architectures on Vercel, this removes the shared-secret problem. (Vercel)

This week's technique: MCP servers. Wire a real tool into Claude so it acts, not just talks.

Most operators have Claude installed and writing code. Few have it connected to an actual tool that does something outside the file system. MCP servers are the fix.

1. Pick a server from the shortlist. n8n-mcp is the fast win: clone the repo, follow the README install, done. 2. Register it in your project scope. Add an entry to .mcp.json at the repo root: a "mcpServers" object containing an "n8n" key, with "command" set to node and "args" pointing at /path/to/n8n-mcp/index.js. 3. Confirm in-session. Open Claude Code, type /mcp. You will see the registered server and its available tools. If the server name appears, it is live. If it does not, the path in your config is wrong, not Claude.

You will know it worked when Claude stops saying "I cannot automate that" and starts saying "I will build the n8n workflow." The tool-call loop closes; the conversation becomes an action.

Claude Code docs: MCP

The move: intercept your API traffic before you scale anything.

Most cost disasters happen in the second week, not the first. You ship the agent, it works, you scale the usage, then the bill arrives and the math is different from what you estimated. tokentap stops that.

1. Install tokentap and run it alongside your Claude Code session or any API-connected script. 2. Watch the real-time terminal dashboard: token rate, cost per call, context window fill, all visible as the session runs. 3. Set a mental threshold before you run any long agentic loop: if cost-per-task exceeds X, the prompt needs tightening or the model tier needs to drop.

You will know it worked when you catch the first runaway context window before it hits your invoice. The dashboard makes invisible cost visible; visible cost is the only kind you can optimize.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. FlowStack, the machine that dreams in pictures. Part three: teaching it to remember a face.

Part two was the bugs that broke the pipeline. This one is quieter, and in a way deeper: getting a machine that paints every frame from scratch to remember what it just drew. It fails in two opposite directions, and both come from the same hole.

The first failure was people who should not exist. One of the channels lived on calm, empty places: still water, open sky, a shoreline with nobody on it. The model could not stand the emptiness. It kept sliding a lone figure into the frame, a face into the clouds, a person onto the empty beach. Left to itself, it will always fill a space. The fix was to stop describing anything human at all, and then to make the prompt insist, in plain terms, on no people. An empty world stays empty only if you demand it every time.

The second failure was the opposite, and harder. In the story channels, one character has to carry a scene across a dozen shots. Same person, three minutes and twelve images later. But the model redraws from nothing each time. Ask it for the same man twelve times and you get twelve different men: he ages, changes his hair, changes his face, between one shot and the next. A story cannot survive a lead who will not hold still.

So we built the machine a way to hold a face. A system that locks a character's likeness and a scene's setting across the whole sequence and carries it forward shot to shot, so the person at the end of the story is still the person from the start, even after the script hands him a haircut halfway through and the cast description has to be quietly rewritten mid-scene.

Underneath both failures is the same thing we hit in lesson one: the machine has no memory of what it just made. It does not know the quiet landscape should stay quiet, or that the face in shot twelve is the face from shot one. Consistency is not something you ask for once. It is something you enforce on every single frame. The work, again, was building the scaffolding that makes it remember.

Next: what all of this added up to, and what it actually looks like when the whole machine finally runs, start to finish, while you sleep.

Fighting a model that forgets what it drew one frame ago? Hit reply and tell us where it keeps losing the thread. We read every one.

Ten repos worth cloning, a studio built from one terminal, and export controls cleared on the models you were waiting on. The gatekeeping keeps falling. Ship something this weekend.

See you Friday.