Step11

What the AI did, what it did not, and where the humans decided

13 October 2026 · 3 min read

aihonestyprocess

A brick-built city under a low golden sun, skyscrapers throwing long shadows.

Steam asks developers to describe how they used generative AI, and on 13 October we sat down to answer those questions properly. This step is that answer, written for players. It is the longest list in the project, and it should be.

The short version: this game is made by one person directing a lot of AI tools, and we want you to know exactly which tools did which job.

What each tool did

Area	Tool	What it did	What a person did
The game's code (about 120,000 lines of C++, about 500 scripts)	Claude (Claude Code, Opus and Sonnet models)	Wrote, tested and fixed it, in roles set up by the owner	Decided what to build, played every build, gave every verdict
LEGO models, maps, sound effects	Claude, through scripts	Models are Python scripts placing LDraw parts; maps and many sound effects are generated by code	Approved or rejected each model from a review board
Story, briefings, in-game text	Claude (a writer role)	Drafted versions and lines	Chose the cast, the plot and the tone; cut and approved everything
Storyboard frames and cutscene video	Grok Imagine	Made the pictures and the video clips from our prompts	Wrote the direction, chose frames, edited the timeline in DaVinci Resolve
Voices in cutscenes and briefings	ElevenLabs	Speaks the lines in voices that are either designed from a text description or standard library voices. None is cloned from a real person	Wrote and chose the lines and delivery; a re-record is planned after picture lock
Interface crests, icons and some images	ChatGPT image generation	Drafted them	Chose and approved them
Some music	Suno	Made some songs	Chose them. The rest is licensed library music
The final logo	Human-approved asset	-	Chosen and approved by the owner. The brand kit on this site is derived from it

Two more honest lines. Early announcer and unit lines used Windows' built-in text-to-speech as placeholders, and we are replacing them. And the game generates nothing with AI while you play: the computer opponent is ordinary scripted code, and the game makes no calls to any AI service.

Where the human decided

A list of calls that were never the AI's:

Every model. Claude proposes a build, the owner says yes or no. We stop iterating where the owner stops.
The story's direction. The two commanders, the feud, the tone. Claude drafted and polished; the owner steered. More in a later step.
The look. Only the LEGO is stylised. That is the owner's line.
Which cutscene clips go in. One take per shot, accepted or redone by the owner's eye.
Tone of the humour. The owner cut jokes that made the villain a joke.
Fun. An AI can run ten thousand scripted games. It cannot tell you whether the sixth minute is boring. The owner and friends play every build.

What the AI did not do

It does not sign in, pay, or accept a licence. The project rule is free assets only; never buy anything, never enter credentials, never click a captcha. Where a licence box needs clicking, a person clicks it.
It does not publish. Builds are uploaded by the owner, not by an agent.
It does not check its own taste. It will happily make forty of something and say all of them look good.

What went wrong

A voice caveat. Speech generated on a free plan is non-commercial. Our notes say to regenerate any line that predates the paid plan. That is a check, not an assumption.
A missing inventory. For the first ten days we did not keep a list of AI-made content. On 25 September a tracker role audited the repository file by file, using each file's own metadata, and produced one. It found items we had forgotten.

What we would do differently

Keep the inventory from day one. One table, one row per AI-made thing, updated when it is made.
Name the tool in the commit. Our commits carry a trailer naming the Claude model that wrote them. We should have done the same for images and audio.