Forty-Nine Agents, Seventy-Four Skills, and a Studio That Asks Before It Builds

The GitHub weekly chart this month has a lot of repositories that wrap Claude Code in a nicer interface. Donchitos/Claude-Code-Game-Studios does something less fashionable: it gives the session an org chart. The repository is a template, not an app. You clone it, open Claude Code at the root, and the assistant inherits a filing system, a set of rules about who may decide what, and 74 slash commands. It picked up around 5,700 stars in a week and sits at 36,845.
The headline numbers are 49 agents, 74 skills, 15 hooks, 14 rules and 62 document templates. Agent counts are the easiest thing to inflate, so the count is the least interesting part. What makes this one worth reading is the part that is hard to fake: a delegation model with an escalation path, and hooks that check whether a subagent actually did what it was told.
Three tiers, and a reason for each
The 49 agents are arranged the way a studio is arranged.
Tier one is three directors: creative-director, technical-director and producer. Tier two is department leads: game-designer, lead-programmer, art-director, audio-director, narrative-director, qa-lead, release-manager and localization-lead. Tier three is 30-odd specialists covering gameplay, engine, AI and network programming, level and economy design, technical art, sound design, writing, world building, UX, prototyping, performance, DevOps, analytics, security, QA, accessibility, live ops and community.
The rules for how they talk to each other are stated in the README, and they are stricter than most multi-agent setups:
Vertical delegation only, directors to leads, leads to specialists. Horizontal consultation is allowed, but same-tier agents cannot make binding decisions across domains. Disagreements escalate to a shared parent, which means creative-director for design arguments and technical-director for technical ones. Cross-department changes go through producer. And no agent edits files outside its domain without explicit delegation.
That last rule is the one that matters in practice. Most multi-agent failures are not reasoning failures, they are scope failures: an agent that was asked to fix a lighting bug rewrites the save system because it noticed something. A written domain boundary, enforced by a lead that has to delegate, is a crude fix and a real one.
The model tiers are a budget, not a hierarchy
Each agent declares a model in its frontmatter. The three directors ask for Opus. lead-programmer and fifteen specialists name Sonnet. The remaining thirty inherit whatever model your session is running.
The README is unusually blunt about this: treat the tiers as seniority of role, not as a billing plan, since most of the studio runs on whatever model you already have. That is the honest version of an agent framework's cost story. In practice you are paying for one session plus a handful of Opus calls on the decisions that shape everything downstream, which is a defensible place to spend.
What the hooks actually check
Fifteen hooks run on session events. Two are worth naming.
log-agent.sh and log-agent-stop.sh write an audit trail when a subagent starts and stops. That is bookkeeping, and cheap.
validate-agent-artifact.sh is the interesting one. It fires when an agent stops and warns if that subagent never wrote the file its brief named. This is a small check aimed at a specific failure that anyone who has run parallel agents has met: the subagent reports success, the orchestrator moves on, and the file does not exist. Catching it at the stop event is the right place, because it is the last moment the parent still has the context to re-dispatch.
The other hooks cover commits, pushes, asset changes, design-doc reads and writes, registries and session lifecycle. Hooks fail gracefully when optional tools are missing, with one deliberate exception: a commit that stages data JSON is blocked when Python is absent, because the validator cannot run.
The requirements list is where the traps are
Setup is four lines in the README and one of them is a warning.
git clone https://github.com/Donchitos/Claude-Code-Game-Studios.git my-game
cd my-game
claude
Then type /start. The system asks where you are, from no idea to existing project, and routes you to a workflow. You can also jump straight to /setup-engine godot 4.7, /brainstorm or /project-stage-detect.
The warning: start Claude Code in the folder that holds CLAUDE.md and project.yaml. Start it in src/ and the project's hooks, deny rules and ask-first permission mode are all skipped. You get an assistant with none of the structure, and nothing tells you.
The dependencies are Git, Claude Code (the native installer is recommended, though npm install -g @anthropic-ai/claude-code works on Node 22+), Python 3, and Bash. Python 3 is not optional in the way it looks: every skill and hook reads project.yaml through it, and without Python every setting silently falls back to its default. Silent fallback is worse than a crash, because the studio keeps running with the wrong configuration and nobody notices until a rule fails to fire. jq is recommended for hook validation.
Primary development is Windows 10 with Git Bash, and the hooks are also tested on macOS with its stock bash 3.2 and on Linux. notify.sh uses PowerShell for Windows toast notifications and does nothing on other platforms, so desktop notifications are Windows-only for now.
Nine engine specialists, one of which you need
The template ships agent sets for Godot 4, Unity and Unreal Engine 5. Each has a lead plus sub-specialists: GDScript, shaders and GDExtension for Godot; DOTS/ECS, shaders and VFX, Addressables and UI Toolkit for Unity; GAS, Blueprints, replication and UMG/CommonUI for Unreal.
Keeping all three in one repository is a reasonable hedge, and it also means the repository carries a lot of text you will never load. The README's own customization section says to delete the agent files you do not need, which is the right advice and also a task nobody does.
The part that is not autonomy
The README states plainly that this is not an autopilot, and the collaboration protocol backs it up. Agents ask before proposing. They present two to four options with tradeoffs. The user decides. Agents show a draft before finalizing. Nothing is written without sign-off. The guided and autonomous modes trade some of those prompts for speed, and the README points at its own setup document for the routes into auto mode.
There is also a test harness for the framework itself. studio-tests/ catalogs all 74 skills and 49 agents with per-category rubrics and behavioral specs, and /skill-test runs a static linter with seven compliance checks, a category rubric, or spec assertions. /skill-improve runs a test, fix, retest loop and reverts if the score does not improve. Testing the framework separately from the game is the discipline that makes editing it survivable.
What this is really for
The README asks its own awkward question in a section title: does it produce games, or documents? A template that ships 62 document templates and a gate for every milestone is betting that solo development fails on process rather than on code. That bet is defensible. The failure modes of an AI-assisted solo project are a hardcoded magic number nobody recorded, a design doc that describes a game three revisions old, and a build that ships with a broken tutorial because nobody ran the smoke check.
Forty-nine agents do not fix those. A path-scoped rule that fires when gameplay code is edited, a /smoke-check that has to run before a build, and a hook that blocks a commit when the validator cannot run have a better chance. The interesting thing here is that someone wrote the org chart down, including the parts where a specialist has to escalate and wait.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.