Last year I had an idea for a tiered avatar system: avatars that evolve as you level up.
The problem was that I was still figuring out image models.
The project doc had two options:
- hire a graphic designer
- use AI generators
It was an AI project, so I went with option two.
The base set
I started with the base avatars, 22 of them. Claude wrote the prompts, I pasted them into Gemini desktop, and it made all 22 in high quality. (This was back when Gemini was generous with images.)
Once I had one set, I thought: okay, what's stopping me from making another? So I designed 2 more sets of about 35 each.
The evolution
With the base models done, I could start on the tiered versions. I call this part the evolution.
I took every single prompt from the first version and made a tier 1 prompt file. Then I wrote the evolution prompts to take tier 1 into tier 2.
While I did that, I had Hermes agent writing the tier 3 evolution prompts at the same time.
Letting it run
Once all the prompts existed, I got a Gemini API developer account and loaded about $50.
Hermes and Claude started generating tier 2 automatically with Nano Banana. I asked for 5 at a time so I could audit them.
We worked like that for the first 10 or so. Then I let the agent just run.
The 20% problem
When I came back, about 80% of the evolutions were good. 20% weren't.
So I brought in Claude to audit the images and refine the prompts for the ones that missed. Then we called the API, ran the script, and generated again.
The result
It took the weekend, but suddenly I had 3 sets of avatars, each with tiered evolutions. It cost $50–60.
If you want to try this
- Do the first set by hand. That way you know what good looks like before you automate.
- Batch small before you let it run. Doing 5 at a time catches bad prompts early.
- Plan for a failure rate. Build an audit-and-regenerate loop instead of trying to make the prompts perfect upfront.