SABLE_SYS — Built with Framer Agents
A choose-your-own-adventure detective story set in a cyberpunk megacity. 22 scenes, four branching paths, three morally distinct endings. Every word, every decision, every pixel — created by AI.
The story, the world, the characters, the branching narrative, the CMS architecture, the UI, the image prompts — everything. No human wrote a line of copy or designed a layout. Three Framer agents were briefed and left to build.
How it was made:
Claude Sonnet 4.6 designed and populated the entire CMS — 22 scenes, a four-path branching node graph, three morally distinct endings, and all narrative copy. 123 credits. One agent turn.
GPT built the initial generative background component — a programmatic pixel art system driven by CMS fields. It didn't survive contact with real Midjourney images, but it proved the visual language worked.
Sonnet then rebuilt the full site UI to a terminal OS aesthetic, bound every CMS field, integrated the images, and made it responsive.
Midjourney generated 22 scene images from prompts written by Sonnet using the CMS data — corruption level, colour mood, and SABLE presence driving every visual.
Total credits used: 2,225 from 10,000. 78% unspent.
Honest account. GPT did real work, it just got outcompeted by a better solution. That's actually a more interesting story — the agents made decisions, iterated, discarded what didn't work. That's not a failure, that's a process.
SABLE_SYS — Built with Framer Agents
A choose-your-own-adventure detective story set in a cyberpunk megacity. 22 scenes, four branching paths, three morally distinct endings. Every word, every decision, every pixel — created by AI.
The story, the world, the characters, the branching narrative, the CMS architecture, the UI, the image prompts — everything. No human wrote a line of copy or designed a layout. Three Framer agents were briefed and left to build.
How it was made:
Claude Sonnet 4.6 designed and populated the entire CMS — 22 scenes, a four-path branching node graph, three morally distinct endings, and all narrative copy. 123 credits. One agent turn.
GPT built the initial generative background component — a programmatic pixel art system driven by CMS fields. It didn't survive contact with real Midjourney images, but it proved the visual language worked.
Sonnet then rebuilt the full site UI to a terminal OS aesthetic, bound every CMS field, integrated the images, and made it responsive.
Midjourney generated 22 scene images from prompts written by Sonnet using the CMS data — corruption level, colour mood, and SABLE presence driving every visual.
Total credits used: 2,225 from 10,000. 78% unspent.
Honest account. GPT did real work, it just got outcompeted by a better solution. That's actually a more interesting story — the agents made decisions, iterated, discarded what didn't work. That's not a failure, that's a process.