Jon's philosophy is simple: give the AI rich context, then let it cook. This little arcade lets you assemble the briefing, tune how controlling you want to be, and see whether your agent ships something elegant or trips over rule fog.
Different missions crave different context. Bug hunts want evidence. Refactors want architecture. Reviews want product intent, because style-only nitpicks are a tragic hobby.
Slide left for command-every-line energy. Slide right for "here's the map, go build something clever."
Click the cards you would hand to the agent. Strong context feeds signal. Doctrine overload feeds theatre.
How much of the mission's real terrain the agent can actually see.
Too many hard bans, commandments, and sacred cows. Very managerial. Not always useful.
Will the agent probably find a clean simplification, or just nervously shuffle lint around?
How likely the agent is to notice the important thing, even when it is annoyingly far from the edited lines.
Pick a mission, choose the context, and press launch. The output here will tell you whether you briefed like a wise collaborator or a jittery committee.
The sweet spot is rarely maximal control. Funny how that keeps happening.
Business rules, product intent, real data clues, concrete examples, and enough freedom for the model to simplify on its own.
Thirty-seven rigid bans, vague mission goals, and a belief that "be perfect" counts as useful engineering input.