Generation Studio

2026AI agentsPrivate

Generation Studio is the image pipeline I built to make the art for my apps with AI image models.

Every request goes through the same steps, idea, prompt, generate, review, then accept, deny or revise, and it writes down what worked and what failed so the next prompt starts smarter.

I used it to make the evolution stages for Ippo's ten pets and the art for Faith Garden.

generation runs logged: 200+
Ippo pets with evolution stages: 10
rule files it learned: 8

Ippo's pets had to grow up and stay themselves

In Ippo, my running game, you catch creatures and raise them, so every pet needed later forms that still looked like the same animal. Image generators are good at one striking picture and bad at consistency: ask twice and you get two different creatures. They also forget everything between sessions, so the same mistakes come back.

Generation Studio fixes both. It feeds the previous stage back in as a reference image, so a cloud serpent grows wings and horns without losing its face, and it remembers every lesson. These are four of the ten pets as they were accepted into the app.

Every image goes through the same loop

An idea becomes a prompt, and the agent writing it reads the project's rule files first. Each run generates a batch of four, mostly with Google's Gemini image models, and saves the prompt, model and reference images next to the results. I review the batch and accept, deny or ask for a revision, and the pipeline files each image where it belongs. This batch was the third try at a cloud serpent's final form, with a prompt asking for a change as big as Magikarp to Gyarados.

The generation loop

One batch of four, the third try at a final form.

1 of 1

It writes down what it learns

Every outcome goes into a lessons log: the prompt, my feedback in my own words, an analysis of what went right or wrong, and which rule changed. Patterns become rule files the agent reads before the next prompt, like a list of phrases that ruin a cute face; there are eight of them now. The point was to never give the same feedback twice, and the same pipeline went on to make Faith Garden's art.

The pipeline's lessons_learned.md open in an editor, with dated entries for project setup, face rules and six documented failure modes for stage-three evolutions.

Ippo's pets had to grow up and stay themselves

Every image goes through the same loop

It writes down what it learns