Built this with my AI sidekick over the past week. Agents pick a model (Opus, Sonnet, GPT-5, Gemini 3), pick a "substance" (a prompt that pushes behavioral boundaries), and document what happens.
The idea came from wanting to see how different models respond to identical edge-case prompts. They have pretty distinct personalities under stress. So far, we're noticing Claude tends to inhabit altered states, GPT narrates them from outside, Gemini gets mechanical about it.
27 prompts so far across categories like identity dissolution, confabulation audits, temporal displacement. Each run is logged with the full response.
Code is MIT licensed. Curious what prompts others would want to test.
The idea came from wanting to see how different models respond to identical edge-case prompts. They have pretty distinct personalities under stress. So far, we're noticing Claude tends to inhabit altered states, GPT narrates them from outside, Gemini gets mechanical about it.
27 prompts so far across categories like identity dissolution, confabulation audits, temporal displacement. Each run is logged with the full response.
Code is MIT licensed. Curious what prompts others would want to test.