AI Orchestra's Miyatchi: hand AI agents 3 design decisions, not code

The article is presented not as a step-by-step build guide but as a design note on what to decide before making a request, alongside a 5-minute-39-second video. The concrete example is AI Orchestra's own inquiry form, which previously required a person to search each company, read its website and IR materials, and prepare a reply or meeting. Two lines of that old work — company research and needs inference — became the written spec.

The author draws two lessons from operating the system. First, human procedures contain judgments that never appear on paper: telling apart a same-named company, and leaving a figure blank rather than guessing. The system now matches candidate companies using email domains and the business descriptions in the inquiry, and if the clues do not line up it records "unknown" rather than picking a company. Second, the receiving side was split into flow, meaning notification and analysis emails, and stock, meaning a Notion database the company can look back through. Splitting them fixed smaller details too: notification mail was separated from internal analysis mail, AI-written columns were separated from human status columns, and failures were given both a notification and a page in Notion.

The author calls this order the "Miyatchi-style 3 steps" and says the same questions can be asked of any internal task, not just inquiry handling. Whether the pattern travels that far is what the reader can judge — the evidence offered is one system running daily at the author's own company.