How I think

01

The hard part is the judgment around the task.

Most founders first look for tasks to hand to AI. The task is usually the easy part. A proposal is hard because someone has to know what this client cares about. A lead is hard to score because someone has to know which signals matter. A client update is hard because someone has to know what needs confidence and what needs care.

In a small firm, that someone is the founder, which is why the work keeps coming back to him.

02

What we write down.

AI already knows the general version of most work. For this firm, it does not know:

• what matters
• what to ignore
• what good looks like
• when to escalate
• when to bend the rule
• when the context changes the answer

Those six are what we write down and give AI as context.

03

Where to find it.

Judgment is hard to describe when you ask for it directly. It shows up in corrections, approvals and exceptions. Three ways I find it:

1. The approval log. For one week, note every piece of work that came back to the founder before it moved: who sent it, what decision they needed, and what the founder knew that they did not.
2. The almost-right test. Take one piece of work the founder corrected. Write down what was technically right, what was still wrong, and the rule someone would need next time.
3. The exception map. Write the normal rule for one recurring process, then every case where the founder would break it. The exceptions are usually where the real judgment is.

04

How we start.

We start with one loop: one piece of work that keeps waiting for the founder. We find where it pauses, what keeps coming back, and what looks done but still needs him. Then we build the smallest thing that carries his judgment, check it against his corrections, and only then pick the next loop.