Show HN: Dreamtap – Make your AI more creative (dreamtap.xyz)

Dreamtap is a lightweight tool/demo (Show HN) designed to combat "mode collapse" — the tendency of large language models to gravitate toward safe, repetitive templates (Claude, for example, often leans on lighthouses and cartographers). Instead of fine‑tuning or changing model weights, Dreamtap injects randomized, unrelated "sources of inspiration" into the model's prompt pipeline before generation. That pre‑generation nudge broadens the model's creative context so outputs diverge from default patterns. Claude can autonomously decide when to pull in Dreamtap’s inspirations; ChatGPT currently requires the tool to be invoked manually. For the AI/ML community, Dreamtap is significant because it demonstrates a simple, prompt‑level approach to improving creative diversity without retraining models. Technically it’s a runtime wrapper that interleaves stochastic context hints with user prompts, preserving the original model while altering its sampling antecedents — a practical pattern for prompt engineering and augmentation. Benefits include richer, less templated storytelling and an easy-to-deploy method across APIs. Trade-offs to watch: potential drops in coherence or reproducibility, shifted safety/behavioral profiles from injected concepts, and the need to tune when and how much inspiration to add. Dreamtap highlights how controlled randomness at the prompt stage can be a powerful lever for increasing model creativity.