Why Do Detailed Prompts Hurt Coding Agents?

David Heinemeier Hansson says the system prompt shipped for Opus 5, and presumably for Fable as well, shrank by 80 percent. Boris, who works on Claude Code, shared that in an interview. The agent needed far less human instruction, and overly prescriptive humans were damaging the output. A hot stretch around agents.md had people micro-optimizing huge files of orders, one instruction after another.

He compares it to a boss who does not know the work and still dictates the method. A programmer sulks and writes worse code when the orders fight better judgment. An agent is no different. The skill he is practicing is to stop over-specifying and get out of the way, leaving a high-level vision. The agile movement in the late 1990s and early 2000s already said that 40 or 50 years of specifying software up front had failed. Nobody knows what they want until they use it. In the agent era, be vague, manifest something, then interact with it. He calls the drop-off the paradox of choice: people can say what they like among a few options almost immediately, and a long list loses them.