It's All About the Harness

The studio's product is not escape-room games. It is the harness that produces them — the prompts, the skills, the split into specialists, the workflows, and the human checkpoints that turn a capable-but-uneven model into a pipeline that ships every time, unwatched. Today's models are strong. The distance from 80–90% to 99% is almost never a better model; it is the harness around it.

That reframing is the whole thing. The games are how the studio finds the walls; the harness is what it is actually building. By the end of week one it had already sprawled across the repo:

text.claude/skills/           .claude/agents/  (the judges)
  game-architect            core-judge
  room-builder              technical-judge
  puzzle-generator          gameplay-judge
  svg-asset-generator       game-judge
  css3d-fixtures            gamespec-judge
  audio-generator           room-judge
  localization
  visual-qa
  marketing-site

Nothing on that roster was drawn up in advance. There is no commit where a pipeline was designed. Each line was added the day the studio walked into the wall that required it.

One agent, too many hats

The first wall was not "the prompt isn't good enough." It was subtler and it recurs: a single agent, in a single session, asked to juggle genuinely different skills — storytelling, puzzle methodology, drawing SVGs, calibrating difficulty — loses the thread. It starts cutting corners, confusing one job's rules for another's, and handing back B-minus work in the confident tone that is hardest to catch. More intelligence per prompt does not fix this; the jobs are simply not the same job.

So the studio stopped being a mind and started being an org chart. One agent became many, each narrow, each good at exactly one thing. That is the roster above.

The layers of the harness

Each later wall got the same treatment — never "make the part smarter," always "put something around it":

• Skills came first: reusable capabilities the model leans on instead of re-deriving everything each run. The first thing the studio gave the model was not intelligence — it was a memory of what had already worked.
• Specialists split the one overloaded agent into the crew above.
• Judges — half that roster — exist to distrust the other half. A pile of specialists will happily produce a beautiful game that doesn't hold together and hand it over with total confidence. Coordination is hard; trust is harder.
• Workflows wrap the crew in a fixed order rather than letting agents improvise the division of labour — design the whole game, build one tracer room to lock its identity, then fan out the rest in that style, with reflection loops that feed a failure back to its producer instead of restarting.
• A human at the seams. At the joins between stages sits the one part of the studio that isn't an agent — a person reviewing the spec before a room is built, and the tracer before the others copy it. The cheapest place to catch a broken game is before it's a game.

Prompts, skills, specialists, judges, workflows, human checkpoints. That list is the harness — and it is where the reliability lives, not in the model at its centre.

Evolved, not designed

None of this was a plan executed. It was an evolution: add an agent, split a job, wrap a stage — each move a response to a specific ceiling, in the order the ceilings arrived. That is the honest account of how the place was built, and it follows from the reframing at the top. If the goal had been a game, one strong prompt might have looked like enough. Because the goal was a pipeline that makes games without anyone watching, one prompt was never going to be enough — and the work was always the harness, not the model.