Zenn developer turns AI prompt rules into incantation

The author starts from a simple frustration: when he asks AI to do hard work, he stops writing one-line requests. Instead he wants to tell the model not to misread his intent, to decide what counts as done, to check the facts, not to settle on the first idea, to question its premises when stuck, to use tools when needed, and to verify the result. When he compressed that into something short, memorable, and still faithful to the mid-task judgments, it came out sounding like a forbidden incantation.

The article's key move is to pull that incantation back into ordinary work. Line by line, the spells map to plain instructions: read the intent behind the words, define the goal and completion condition first, go back to primary records instead of stopping at second-hand information, check dates and versions, separate fact from inference and hypothesis, keep multiple candidates, look for evidence against your own view, set comparison criteria, use search or tools when knowledge is missing, re-examine assumptions when stuck, respect constraints and prohibitions, create-test-break-fix, leave uncertainty visible, stop when enough is enough, and ask whether the answer actually works.

The author notes this is not far from official prompt design guidance from OpenAI and Anthropic, which recommends giving high-level instructions, organizing roles, examples and context, breaking complex tasks into clear steps, and giving strong guidance at the right level of abstraction. The genuine limit is that short spells drop specifics — numbers, API specs, permissions, exceptions, file names, and verification conditions — so the author suggests three layers: incantation as the human- and AI-friendly entrance, a work contract for purpose, inputs, outputs, and stopping conditions, and a detailed spec for exceptions, numbers, permissions, and tools. He also raises an open question of whether another AI shown only the spell could reverse-engineer the original workflow, but he treats that as a hypothesis to test across models rather than a result.