DreamFlect

Dreamflect — a self-reflection app that doesn’t interpret your dreams for you; it helps you interpret them yourself.

You type or speak one dream. From there, Claude Opus 4.7 never writes a sentence — every screen is built at runtime by the agent calling frontend tools: a gap-fill question, an archetype card, an archetype-specific probe, an interpretive frame, and a shareable takeaway rendered as an A2UI surface layered over AI-generated dream art (gpt-image-1).

Built on CopilotKit v2 (headless useFrontendTool, single-route runtime), A2UI v0.9, Next.js 15, Better Auth, and Neon Postgres. Each tool’s handler returns a Promise that pauses the agent loop until the user clicks — the interaction is the tool result, so the model literally cannot advance until the UI responds. No chat bubbles. No assistant text. The interface IS the conversation.

Prior to the hackathon: drafted mid-fi wireframes in Claude Design (exported to dreamflect-handoff/ as read-only reference — re-implemented from scratch in shadcn + Tailwind during the hackathon) and brainstormed the user flow with Claude — the core philosophy, the morning/evening rhythm, and the ~18-archetype taxonomy with its archetype → probe-variable mapping (chase → agency, falling → impact/witness).

Built during the hackathon: all code — the CopilotKit v2 + A2UI v0.9 generative-UI architecture, Promise-pausing tool handlers, system prompt, Whisper voice capture, gpt-image-1 dream art, next/og share routes, Better Auth + Neon persistence, memory loop, and evening-reflection mode.