Building Is Believing

01 — Inside the Box

What is inside an agent?

The word is everywhere. So let's open it up.

YouTube | OpenAI
A two-minute look at one in motion — a third-party risk management agent: Workspace agents in ChatGPT.

An agent is mostly a specification — written in plain English. You describe the parts, and the system assembles the behavior. Here is the anatomy. Seven parts.

Role

Who it is, and the judgment it stands in for.

Tools

The actions it can take — search, calculate, draft, send.

Skills

The know-how it brings to a task, packaged and reusable.

Files

The context it reads — your documents, your numbers, your notes.

Sources & Tools

Where it is allowed to look, and where it is allowed to act.

Operating rules

The guardrails, stated plainly — what to do, what to leave alone.

Output

The shape of what comes back — a memo, a table, a decision.

---
name: policy-comparison
description: Compare two insurance policy forms (expiring vs.
  proposed) and return a ranked, cited list of material
  coverage differences. Use when two policies or quotes are
  supplied for comparison.
---

## Method
1. Normalize — parse each form into one skeleton: declarations,
   definitions, insuring agreements, conditions, exclusions,
   endorsements, limits / sublimits, retentions.
2. Align — match section to section across both forms; resolve
   cross-references (an endorsement that amends a definition).
3. Score materiality — reduced limit, new or broadened exclusion,
   narrowed definition, added condition or warranty, changed
   retention → rank High / Medium / Low.
4. Cite — clause + page for every difference; quote the operative
   words; never infer coverage; flag ambiguity to verify.

## Output
A side-by-side table by section, a ranked list of material
changes, and a short "what to ask the carrier" list.

## Files (bundled, read as needed)
exclusion-taxonomy.md · defined-terms-glossary.md ·
materiality-checklist.md · output-template.md

02 — Inside the Conviction

Conviction is a self-belief

Curiosity asks the question, and is willing to go down the rabbit hole. Persuasion allows another voice to move your position. Conviction is the slow, firm certainty you build for yourself. Which is why conviction grows from the inside.

So how does anyone arrive at conviction with AI? Two paths worked for me.

Building is believing

Thoughts to prototype is a matter of hours and days, not months and years. When an idea becomes a thing I can touch, I stop debating whether it is possible and start leaning into what it is. So build something useful. The making is the believing.

Comprehension feeds conviction

To trust a tool, I want to see how it is built. In the simplest form, the building blocks of AI are skill and constitution. Skill is talent — what the system can do. Constitution is personality — how it chooses to behave. To understand them, I used AI to draw the maps myself.

Hope this helps you too.

03 — The Uncomfortable Half

The job is uncomfortable

It is comfortable to dwell on the upside. It is uncomfortable to sit with the downside. And sitting with the downside is the job description of a risk manager — isn't it?

There are many forces leaning into what AI can do. Fewer people, mostly researchers, are doing the patient work of asking how it fails. That work can also be shared with risk managers. The risk managers of the AI era want to understand the risks of AI, and the good material is available:

04 — The Honest Answer

"What can AI do for me?"

Everyone is asking it. The honest answer is "it depends" — on who is doing the asking. A claims lead, an underwriter, a risk advisor, a placement broker, and a CFO each need a different answer.

So here is a possible next move: build the agent that helps deliver the suitable answer for each person on the team. Hand it four things:

"Here is my role.
Here is what I do.
Here is what filled my last 90 days.
Where can AI help?"

The last 90 days matter most. Your calendar is an honest record of where your hours actually go. Feed that in, and the question stops being abstract. It becomes your question.

05 — Try It Now

Build a prompt that works

"I want to build, but I don't know how to prompt." Fair. So let this do it with you.

An effective prompt has its own anatomy — the same idea as the agent, scaled to a single ask. Name these six parts and most of the work is done:

• Role — who it should be
• Context — what it's working with
• Goal — what you want
• Task — what to do
• Rules — what to honor
• Output — the shape of the answer

AI does two jobs especially well: it compresses the research, and it surfaces hidden information. Answer a few questions below, and it assembles a prompt you can paste into your capable AI.