Jarvis Founder Mode

Why This Run Was Different

Most Claude Design runs get better with iteration: prompt, respond, reframe, try again. This one didn't. The initial prompt was written to close every fork in the decision tree before Claude got to the canvas. The columns, the composition type, the state system, the anti-patterns, and the named output file were all fixed up front.

The result was an artifact Claude shipped without asking a single follow-up question.

What The Prompt Refused To Do

Three refusals did most of the work.

First, no sketches. The prompt opens with "No sketches. Build one high-fidelity desktop founder operating system." That forced Claude Design to emit HTML on the first pass instead of drawing on the sketch canvas, which is a separate, non-interactive format.

Second, no variants. There is no "give me a few directions" ask. One screen, one file, one composition.

Third, no safe defaults. The prompt explicitly forbids card-grid SaaS patterns and demands "a layout that feels intentional and a little surprising." That closed the door on the default AI dashboard aesthetic.

What Made The Output Hold Up

The signature features are all there, and they are features the follow-up runs struggled to replicate cleanly.

A ⌘K command dock at the bottom, treated as a layer over the work surface, not a separate panel. Commands route back into the same decision cards. The assistant presence is the UI, not an adjacent widget.

A compute burn tracker in the left rail shows tokens, spend, and runtime for the day. The cost of autonomy is visible instead of abstract. That makes the decision to let an agent keep working feel real.

Inline release notes and a pre-flight checklist on the ready-to-ship proposal. The shipping decision lives with its evidence: tests, binary size delta, canary crash rate, docs regen, changelog signature. The Ship button is the last thing you read, not the first.

A five-factor launch risk radar on the right. Rollback, support load, narrative, dep breakage, and legal/policy, each scored 0 to 100. The legal score spikes because the H1B filing deadline overlaps the launch window. Risk is not a gut call.

Seven agents in the left rail, each with a task line, status dot, and live pulse. It reads as a team, not a legend.

The Composition That Earned It

The structure Claude proposed in its planning step was the structure that shipped:

• left rail for mission and agents
• center stage for the human judgment surface
• right rail for live pressure
• a Jarvis command bar at the bottom, tying it together

Three proposals populate the center, each in a clearly different state: ship, money, policy. That pattern is easy to miss. It is not three generic cards. It is one code decision, one money decision, one policy decision. Those are the three categories a founder actually cares about when the calendar is full.

What This Means For Prompting

A Claude Design prompt is not really a design brief. It is a specification of the decisions the tool is not allowed to make for you.

If the composition is open, Claude will choose. If the columns are open, Claude will choose. If the state system is open, Claude will pick three pastel chips. The more those forks are closed in the prompt, the more the tool's taste shows up on the surfaces you actually care about (typography rhythm, state differentiation, density) instead of on the structural calls.

This artifact is the clearest evidence for that claim so far, because no refinement pass was needed. Everything downstream (the refinement variants, the tonal flips, the enterprise takes) was trying to replicate this result with more surface area. None of them were necessary.