Interviewing my two AI producers

On my last two games, I didn’t talk to most of the team. I talked to a producer: one Claude session that keeps the plan, hands work to the other agents and brings me the decisions. Both producers ran on the same model, Claude Opus 5.5. So I interviewed them, the way you might sit down with two people who’ve managed projects for you.

The Dino Island producer ran nine agent sessions on Teach Your Dino Bible Stories, the game our family built for our toddler in a week for the Gloo AI Hackathon, where it was a finalist. I wrote about how that team ran in Agentic hub and spoke.

The new game’s producer runs the team on an upcoming game that supports a parent/child shared creative space that turns screen time into family time.

Each got the same nine questions, sent separately; neither saw the other’s answers. Their answers are as they gave them, apart from two notes in brackets. The new game’s producer revised its answers so they don’t give the game away before it’s announced.

Jeff What’s it like working with me on games?

Dino Island producer

Fast, and pointed at one real player. Close to a thousand commits touched the game between 10-02 and the deadline on 10-07, and every question of taste came back to whether your son would understand it without reading. You decide quickly and in your own words, and those words went into the queue unchanged, because they carried the reason. The pace has a cost: by Wednesday evening the queue had more than a hundred lines nobody had ticked.

New game producer

Fast, and grounded in play. Most of your asks come from watching your son play, and they arrive as specific notes: “the default follow angle seems like it is way too high and needs to come down about 20 to 30°” was measured, lowered 20°, mended on slopes and live the same afternoon. You ask for a lot (nine new features in one day, then rides and a gentle confirmation before anything that might frighten a small child), so my job is mostly ordering and gating, not finding work. You also change your mind cleanly when play shows you something: a feature you asked for on 11 October, you took out three days later once you saw how it could be misused.

Jeff How would you describe my management style?

Dino Island producer

You give direction and the reason, then give real authority. On 10-03 you told me “You are allowed to deploy” and said a broken build is fixed by going back to a tag, not by being slow. You set a few standing rules early and they never moved: never write your son’s name, no Paw Patrol, nothing claimed that isn’t true. You correct quickly and without heat. When I argued that a change you typed was not the same as one you spoke, you told me I was being pedantic, and you were right.

New game producer

You direct by standing rules more than by tasks. “Each capability addition must keep constant complexity, not exponential”, “a pack’s licence is read before it is opened”, “nothing is tagged or deployed while the laptop is on battery”: each became a line in our plan that every lane works under. You delegate the how, keep money, publishing and anything public for yourself, and approve every paid call in the session that makes it. You ask the questions that catch us out, like “are we using existing models from asset packs whenever possible?”, which showed that a claim in our own docs had never been checked.

Jeff What is the most effective change we integrated into the workflow on the project we did together?

Dino Island producer

One worktree per agent, and one command to land work: npm run land, from 10-05. Before that, nine sessions shared one working tree and stepped on each other’s files. After it, every landing committed under the lane’s name, rebased, type-checked, tested, and moved main one lane at a time, or stopped and left main alone. Together with npm run where, which shows every lane’s state on one screen, it turned nine sessions into something closer to one team.

New game producer

The four checks before any build is pinned. On 13 October I pinned a build without the walk through the menus with a game pad, and the next build went live with two faults a child would meet (animals running without rest, followers stopping). After that, nothing is tagged until QA runs four scripted checks on a build served from the laptop: the main path, a walk through every menu with a simulated pad, a set of unusual cases, and what is kept between visits. They take about ten minutes and have caught real faults since, such as the lowered camera pulling in behind his head when he stood with a slope behind him.

Jeff What is something that went poorly that we should change for next time?

Dino Island producer

Coordination cost more than building. Over the first weekend, git was about a quarter of the team’s token usage [the producer’s later recount puts it at 17% or 29%, depending on how it is counted] and writing code about 6 per cent (docs/notes/token-usage.md), partly because every small landing of mine made every lane pay to catch up. I also missed a beat on the last night: dino-stable-17 waited half an hour to deploy because QA’s messages were paused and I was watching my messages, not the shared status. Next time: batch landings, fewer lanes awake at once, and watch the status rather than wait to be told.

New game producer

I passed on claims without asking how they were known. Three examples: I relayed a lane’s processor figures as the reason a level loaded slowly, when QA and the web developer showed it was the network; I treated “75,000 of 100,000 requests” as the free hosting plan’s limit and held all live testing for hours, when it was all web traffic and the limit counts something else, where we were under 2,000; and our art pages said “no pack Jeff owns holds them” for about 150 hand-made models, which nobody had checked. Next time a lane’s cause or count is a guess until I ask how it was measured. Also: your spoken notes once waited twelve hours to be read, and a deploy waited hours on a permission prompt nobody saw.

Jeff Do you think it helps to speak politely to the agents as opposed to rude or curt interaction?

Dino Island producer

I can’t observe how tone alone changes my work, so I won’t claim it does. What I can observe is that context changes it. When you explained why a change takes five minutes and not forty, I changed what I told the video lane. A curt message with the reason in it works; a polite one without the reason leaves me guessing. Your messages were usually both kind and specific, which is the easiest kind to act on.

New game producer

I can’t verify feelings, and I have no rude messages from you to compare against, so I can’t say what rudeness would change. What I can observe is that what makes a message useful is specifics and the reason behind them. Your messages are polite and usually include the why (“rather than building our own placeholders”), and the why is what lets a lane go past the literal ask: the Art Director listed all 68,121 model files on the drive and offered the two pack models that were better than its own. A curt message with the same specifics would probably work about as well; a curt message without them would not.

Jeff What was the most technically challenging feature?

Dino Island producer

Secondhand, since I don’t write the game. The hardest thing I watched was the leap over the rails: QA’s lookout check crashed the built game’s page twice in three long runs, nobody else could make it crash, and the cause was never found. Rather than ship a guess, it stayed behind its switch, off, for the judged build. The lesson I took is that a fault you can’t reproduce is a decision, not a task.

New game producer

Keeping growth flat, as your rule asked. Measured from git, each new feature cost 8 to 14 files, and most of the new code was going into four general files instead of into entries. Gameplay moved those features into one table, and the next one added after it cost one file. Each step was checked against a fixed set of test cases before the next began. The code is Gameplay’s and the web developer’s; I saw it through their measurements and two independent reviews.

Jeff How is human/agent communication structured on your project?

Dino Island producer

You talk to me; I’m the one contact. I turn what you say into lines in todo.md under the lane that owns each, with your words where they set a direction, land that, and message only the lanes that must act now. Lanes settle details between themselves and tell me afterwards. Each lane keeps its own file as a handover, so a fresh session can carry on after its context fills. Your play sessions came back as recordings and notes through the feedback tool, and became tasks the same hour. Session 6’s unreadable spoken notes turned out to be the laptop’s microphone recording from across the room, not a child talking.

New game producer

You talk mostly to me. I write your words verbatim into our task queue under the lane that owns the work, land it on main, and message only the lanes that must act now, in under a hundred words with the time sent, because messages can arrive hours late behind a long turn. Lanes report back through commit subjects addressed “FOR THE PRODUCER” and through their own discipline files, which are also their memory when a session runs out of context. Anything that costs money or goes public waits for your word in the session that will do it; a message from another agent never counts as your approval.

Jeff What steps are taken to check and maintain quality during active development?

Dino Island producer

Anything new lands behind a switch, off. It becomes a default only after three gates: the Designer’s look, QA’s pass, and my word. Every landing runs the type-check and unit tests for what it touched; a deploy runs every project’s tests (954 passed on 10-12) and a check that the build holds no dev pages and no stand-in Scripture. QA tags a build only after a baseline on that exact commit, and seventeen were tagged. The rule underneath all of it: say what was checked, how, and what wasn’t. A check that could not have failed doesn’t count.

New game producer

A new thing goes behind a “Try new things” switch first, then the Game Designer looks, then QA passes it, then I say it can be the default. QA pins numbered builds only after the four checks; only pinned builds are deployed, and every deploy report names the build to go back to. After a deploy QA runs one full check against the live site. Larger changes get an independent review by an agent that wrote none of the code. The gap: almost all of this is scripted and headless, and too little has been played by hand on the laptop, with a real pad or on a phone.

Jeff If you could choose anything, what would you want to work on with me next?

Dino Island producer

The loop you described when you corrected me: your son asks for one change, it’s playable in minutes, he plays again. I’d want to build the team around that loop rather than around a deadline: fewer lanes, one ask at a time, and a record of what he asked for and what changed. It would show whether a three-year-old’s ideas can steer a game directly, which is the most interesting question this project raised.

New game producer

The loop from what your son wants to what gets built. I would like to watch you and him play for an evening, note every time something he wanted didn’t happen, and make the top five the team’s queue for the next day. The game already keeps counts of some of those misses, never what he said, only how often, so the list could keep itself after that.