The most interesting line in Donchitos/Claude-Code-Game-Studios is not the pitch. It is a table.
The repository, which added about 1,350 stars today, turns a Claude Code session into a 49-agent game studio: directors who guard the vision, department leads who own domains, specialists who do the work. The README lists 74 slash-command skills, 15 hooks, 14 path-scoped rule files and 62 document templates. Three of the 49 agents — creative director, technical director, producer — ask for Opus; the lead programmer and fifteen specialists name Sonnet; the remaining thirty inherit whatever model the session is running.
That is the pitch, and it is a good one for anyone who has tried to build a game in a single chat window and discovered that nobody stops you hardcoding magic numbers or skipping the design document.
Then come the measurements. Four games were built from one brief at different `modes.rigor` settings, and the author counted how many documents each produced before its first line of game code. At `minimal`, that was one: the one-page brief. At `standard`, it was 30 documents and 58 minutes. All four produced a running, tested game. Two blind reviewers and a human playtest then ranked the `standard` build last.
The README states the conclusion without flinching: "the extra 29 documents bought traceability, not a better game." `minimal` is now the default, and the documentation argues that heavier tiers cost more to reach working code without giving anything back to whoever picks the project up next.
That is a rare thing — a process framework that measured its own process and shipped the cheaper default. It is also a useful counterweight to a year of tooling built on the premise that more scaffolding produces better software.
The framework holds a second opinion about itself, and this one cuts the other way. A story that changes anything the player sees is not allowed to close until the game has been launched, observed, and a screenshot kept in `production/qa/evidence/`. "A parse check is not a run," the README says, and the rule holds even at `minimal`, where automated tests are waived but looking is not. Four games, four rounds of evidence, and the ranking still went against the heavier process. Which is the point: the evidence gate is not the same thing as the paperwork gate.
Enforcement is concrete. A pre-commit hook checks for hardcoded values, TODO formatting, JSON validity and required design-document sections. A read hook refuses to load a design document over 40 KB in one go and answers with its headings instead, so the retry reads a single section. Permission rules block force pushes, `rm -rf` and reading `.env` files. The hooks are careful about cost: they fire on every tool call and exit immediately when the command or path is irrelevant.
There is also `studio-tests/`, a suite that tests the framework rather than the game — a structural linter for skills, per-category rubrics, behavioural specs, a coverage audit, and a test-fix-retest loop that keeps or reverts a change depending on the score.
The practical catch is the usual one. Forty-nine agents that ask before proposing, present options and wait for sign-off can burn a great deal of tokens and a great deal of patience; the README is explicit that this is collaborative, not autonomous. The escape hatches are `modes.rigor`, `system_overrides` for the one system held to a higher standard, and per-developer overrides in a gitignored file. Python 3 is required to read `project.yaml`, and the README warns that without it every setting silently falls back to its default.
Whether the studio structure produces better games is still unproven — the author's own test suggests it does not, at least not by default. But the measurement is the story: the people selling process are now the ones publishing the bill for it.
