Data-Driven Prototyping and Structured Midjourney Prompts for Elite Results

By: Ravi Mehta Host: Claire Vo Source: How I AI (ChatPRD / Lenny’s Podcast Network) Type: podcast

Summary

Two workflows addressing the “vibe prototyping” problem — asking AI to simultaneously handle UX, code architecture, and data structure produces mediocre results. Workflow 1 (Data-Driven Prototyping): Generate structured JSON data with Claude first, using Unsplash MCP server for real image URLs, then feed that JSON to a prototyping tool (Reforge Build) with a minimal prompt. Separation of concerns lets the AI focus purely on optimal UX for the given data. Enables easy iteration: swap JSON for different destinations/languages, stress-test with edge-case content lengths, augment existing data without starting over. Workflow 2 (Structured Midjourney Prompting): Subject-Setting-Style framework — define what you’re showing, where/lighting, and how it should look using photographic vocabulary. Film stocks (Kodak Trix, Fuji Color C200, Fujifilm Provia) and camera metadata (Leica, 50mm, F1.2) steer the model toward higher-quality aesthetic regions of its training data.

Key Ideas Extracted

Separation of concerns in prototyping: Don’t ask AI to solve UX, data schema, and code architecture simultaneously — generate structured data first, then let the prototyping tool focus solely on UI/UX
MCP servers for real media: Unsplash MCP in Claude provides actual image URLs, eliminating hallucinated/broken images that plague vibe prototyping
JSON-first prototyping enables iteration: Swap datasets for different scenarios (destinations, languages, content lengths) without rebuilding the prototype; stress-test edge cases in design reviews
Subject-Setting-Style framework for Midjourney: Three-element prompt structure that produces professional-quality images vs. generic “vibe prompt” results
Film stock + camera metadata as “cheat code”: Specific photographic references (Kodak Trix for B&W contrast, Leica for premium aesthetics, F1.2 for bokeh) point the model toward higher-quality training data regions
Art literacy as AI skill: Understanding visual vocabulary (lighting, composition, style references) directly improves prompt quality — recommend using Claude/ChatGPT to describe photos you like to build this vocabulary
“Elite” keyword for quality guidance: Adding qualifiers like “elite” to prompts guides AI toward higher-quality training data associations
AI enables “nice-to-have” features: Personalization and delight features that get cut from scope become feasible with AI — competitive advantage for consumer products

Notes

Published Sep 28, 2025 on ChatPRD blog. 10-min read.
Tools: Claude (JSON generation), Unsplash MCP server, Reforge Build (prototyping), Midjourney (image generation)
Ravi also has an AI Strategy course with Brian Balfour through Reforge
Practical detail: Claude generates JSON with real Unsplash URLs via MCP calls, but this takes extra time due to multiple API calls
The data-driven approach produces clean, componentized code where data lives in lib/sampleData.ts for easy swapping
Sponsors: Google Gemini, Persona

Raw Content

Re-scraped from ChatPRD 2026-02-15. Full article content captured in Summary and Key Ideas above.