Prompt Builder: A Workbench That Turns Visual Intent into Model-Correct Prompts

I used @hyperagentapp by @airtable to finally finish Prompt Builder. It’s now live: a browser-based workbench that turns visual intent into model-correct prompts for Midjourney V7, FLUX.2, Nano Banana, and GPT Image 2. I had already built the core idea. The unfinished part was the hard last mile: the visual system, the Agent workflow, QA, PWA caching, and making sure the deployed site actually matched the code. I used Hyperagent as the operating layer for finishing the project — not as a one-shot chatbot. An art-direction agent developed distinct visual directions from the real product. A coding agent implemented the chosen direction. An independent acceptance agent reviewed the result, so the builder never graded its own work. A persistent LOOP-STATE document carried decisions, defects, and next actions between runs. Browser checks verified the actual live site. A green GitHub push did not count as shipped. I made the creative call: “One-Sheet” — coated-paper white, ink black, one Klein-blue accent, zero rounded corners, and the hierarchy of a printed production document. What shipped: - 100+ visual presets - Four model-specific output formats - Manual controls that work without an AI key - An optional BYOK conversational Agent - English and Chinese input - No account and no backend - One 357,539-byte HTML app The creative decisions stayed mine. Hyperagent let me run the work like a small production crew: clear roles, real handoffs, persistent state, independent review, and a finish line. Born on a film-prep desk. Built for anyone who prompts. Where do your projects usually stall: the core build, or the last 10% required to make them truly finished? #AIAgents #BuildInPublic Try Prompt Builder: https://t.co/UGRy8aPZdD Open source: https://t.co/K5jTihAuOA