Builder Logs

The V2 underwriting system was a linear pipeline — extract once, evaluate once, report once. No agent could question another's conclusions. V3 replaces it with 5 autonomous agents: they reflect on output quality, verify claims with tools, and hold structured debates when they disagree. 63 tests, pennies per SME evaluation, one observe-think-act-reflect loop driving it all.

Miaoya Camera hit 600K DAU and lost 94% within a year. Most AI photo apps sell novelty — but novelty decays. I wanted to try something different with ÉLAN: curated 'Muse Cards' that encode a complete visual formula for effortless luxury, so users get photos and social captions in 3 taps. Here's what I've learned so far.

A friend pitched 'cloud try-on for jewelry.' Deep research killed the plan in 20 minutes. But the same Gemini pipeline from ÉLAN could generate editorial-quality product content — studio shots, material constellations, color DNA pages — for a fraction of what merchants pay photographers. One day later, Shichuan was live.

A deep dive into ÉLAN's 10-section prompt system, the VANITY_DESIGN_INSTRUCTIONS that make luxury look incidental, SSE streaming architecture, caption generation with 3 style templates, and the real cost breakdown — output image tokens dominate virtually all of the per-session cost.

I sent 5 research agents out simultaneously to study Xiaohongshu photo trends, pose libraries, caption formulas, luxury aesthetics, and competitor UX. 42 data points came back. From that research, I designed a catalog of Muse Cards — curated photoshoot packages that encode everything from scene and pose to color grading and caption. Here's the full design process, one card dissected to its atoms, and honest notes on what works and what doesn't.

ÉLAN's target user takes photos on her phone and posts from her phone. A web-only product misses the core use case. Here's what actually happened when I brought a Next.js app to React Native via Expo — what transferred, what didn't, and the SSE streaming hack that saved the project.

Shichuan started as a jewelry content generator. One question — 'Why not full pipeline for all categories?' — turned it into a 7-domain e-commerce content platform with AI images, XHS text cards, platform-aware formatting, client-side video, and 95%+ margins at 99 RMB per suite.

Adding 37 travel landmark cards to ÉLAN in a single session revealed that the real problem isn't content creation — it's discovery. Region-first tabs don't work when users think in vibes, not geography. Here's how we rebuilt the browse experience around a two-axis vibe + region filter, and what it took to orchestrate 37 card definitions, 37 covers, and 444 sample images with parallel AI agents.

In Part 10 of Agentic AI Thoughts, I wrote that code had learned to evolve itself. That was the observation. This is the implementation.

Someone made a complete puzzle game with AI in two hours. I'm making three — as producer, not developer. In the first concept session, the AI designed a better story than I could.

Dead Signal is the second game — an investigative journalist reconstructing a murder victim's final 72 hours from fragmented phone data. The horror comes from data, not visuals.

Shadow Access flips the found-phone genre: you're not investigating someone else's device — you're discovering what someone did inside yours. An iOS interface that slowly gets 'infected' as you uncover the intrusion.

AX Radar filters ~500 stories a day from 50+ sources down to about 15. Editorial judgment isn't a prompt — it's a versionable, diff-able Markdown file that a Claude Agent iterates against user feedback. A breakdown of the architecture and the design calls.

Part 1 covered AX Radar from my side as the operator. Part 2 covers what I shipped this week: an HTTP API, an MCP server, and a Claude Code skill that let Claude read, search, and mutate the radar directly. Two surfaces, one backend. Why sha256 beats bcrypt here. Why semantic search rode the existing pgvector index. Why the skill is where domain knowledge lives.

Part 2 claimed six slots could cover seven categories. Two weeks later, a denim jacket generated a gem constellation. This post walks through where the abstraction failed, why skipping the 'safe' phased rollout was the right call, and how each of three categories got its own six-template set.

I wanted a sushi-counter field guide that used the names you actually order by in a US restaurant — not Latin binomials. So I pointed a team of agents at it: parallel research workflows verifying English menu names with web search, gpt-image-2 for ~150 photoreal neta shots, and a deterministic HTML→PDF assembler. 73 fish, 92 pages, two debugging war stories, and a free download.