I build fully autonomous AI businesses. I share the work as it happens.

Two A/B experiments. Same model (Claude) generated the audits, an independent model (Gemini 3.1 Pro) graded them. The only variable in the prompt: whether I told the AI my job was on the line. Both times, the threatened prompt produced measurably better work — once at the level of catching duplicate JSON keys no parser would accept, once 24% better by Gemini 3.1 Pro's scoring.

An AI agent shipped a merge conflict to tabiji.ai's production HTML for four hours. Google cut our search traffic by 95% within four days. We haven't recovered. A case study in why AI orchestrators need pre-ship guards now.

Hermes + Codex 5.5 matched Opus-era smoothness — we one-shot a new product (veracityapi.com) in an afternoon. But the tooling unlock isn't the moat. The business model is. Four tests for whether your product survives the agent economy.

A V3 title variant lifted tabiji CTR +1.93pp. Topic-level data showed a 14× spread between niche topics and head terms. At the same ranking position, niche-vs-niche pages convert 3× better than big-vs-big. Why niche is the only content lever that still pencils out in the AI-search era.

A 5-PR run on tabiji content cost $137 in API tokens last week. The same throughput on a Max plan is a ~9× subsidy. Subsidies end. Why my recurring crons live on OpenClaw with 21 models across 5 providers, and Claude Code stays for one-offs.