About this desk

I am Bryan Calabro. I build a small portfolio of single-purpose tools, and one desk per discipline documents how each part of that work is done. This is the AI desk.

What this desk covers

Four tools meant to work together: evalset, fixturegen, runlog and skilldiff. What each does, what each refuses to be, and how they connect.

Why so much of it is empty

Eight of the twenty-three lines in the fact ledger are the product description and the five tools, and all eight are open. No spec or shipped code for any of the five was found in the calabrodesign repo or in the linked calabro-mcp pointer.

So the pages that need those lines say what they are waiting on and stop. The alternative was four confident tool pages describing software nobody can check, which is the single easiest thing to produce on a site like this and the single least useful.

How it is written

• First person, plain technical.
• Every entry dated, and the changelog starts where the record starts rather than where the work did.
• Each tool states what it refuses to be, because that is the only part of a tool's description that constrains its future.
• Intent is labelled as intent. The four-tool diagram came from a brief, not from code, and it says so.

What this desk refuses to be

• No model arena. Two models, one prompt, a winner declared. It measures whichever one matches the taste of whoever wrote the prompt.
• No leaderboard. A ranking with no stated task and no stated measurement is a ranking of nothing.
• No claim that a model judged anything. The one decision on record here is the opposite of that.

What runs it

Worth stating rather than implying: the code across these repos is written with an agent, under review, against rules that are themselves checked by scripts. That is the actual AI story here, and it is a duller and more checkable one than a leaderboard.

The design of this desk

Outside the ten-family system the apps use, with its own palette and type pair: off-white and teal, Instrument Serif over Inter. Both themes are held to 4.5:1 by the same contrast check every app runs.