The First Screen Is a Control Surface
mission:
help ryan turn real work into clear writing that brings in clients
progress to goal:
portfolio [####......] 2/6 tasks
inbound [###.......] 2/8 tasks
partner pipeline [#####.....] loaded
priorities:
01 turn the strongest buildlog item into public proof
02 keep source-sensitive claims tied to evidence
03 leave other agents' domains alone
The old launch surface was technically correct and operationally weak.
It could prove the agent had context. It could show that a boot compiler ran. It could carry identity, rules, and recent state into the model. But it still asked the human to do too much work at the exact moment when the system should be reducing work.
A session should not open with a payload. It should open with a readout.
The Product Mistake
The first version of the opening brief treated startup as a context problem. That is reasonable if the audience is the model. It is wrong if the audience is the operator.
Ryan does not need to see every source file the agent was given. He needs to know three things fast: what this agent owns, how the owned goals are doing, and what work is first in line.
Counts alone were not enough. A line that says an agent has six tasks still makes the operator ask, "Which six?" A card that shows the top three priorities changes the launch from a status report into a decision surface.
What Changed
The renderer now uses an 80-column instrument layout by default. It has the agent and company in the header, a plain mission section, progress rows with block bars, and numbered priorities. The older plain render still exists for debugging, but the visible default is the selected product surface.
That sounds like presentation work. It is more than that. The card forces the harness to decide what deserves first-screen attention.
| Before | After |
|---|---|
| Raw boot context proves the agent has memory. | The card shows the operator what matters now. |
| Task counts require a follow-up question. | Top priorities are visible at launch. |
| Screenshot mode was optional. | The selected instrument is the runtime default. |
Why This Belongs in the Harness
The Boot Card is small, but the pattern is large. Good AI systems do not just add memory. They turn memory into a working surface.
A founder does not want a model that can recite every meeting note. A founder wants the right pressure, the right next decision, and the right boundary in view before work starts.
That is the same pattern behind the rest of the Gauges Green harness: compile context, route attention, attach gates, and show the state of the work before judgment begins.
The Evidence
The build was not accepted as a vibe. The final pass kept the frame to 80 columns, removed the old visible "Boot Card" label, checked four agent dry-runs, and passed the focused opening-brief test suite.
The useful lesson is simple: if the first screen is the first decision surface, it has to be designed like product, not printed like logs.