Guide

How to build your own AI review board with ChatGPT, Gemini and Claude

Before anything important leaves our shop, my agent sends it to reviewers on three AI models from three different companies. Each one checks it alone and comes back with a verdict and a list of what has to change. You can run the same setup by hand today with the chat apps you already have. No code required.

If you want the story of why we do this and how it runs inside our business, read How we built an AI board that checks our biggest calls. This page is the how-to.

What you need

• Accounts on two or three AI chat apps from different companies. For example ChatGPT (OpenAI), Gemini (Google) and Claude (Anthropic). Two is the minimum. Three is better.
• The piece of work you want checked: an email, a quote, a contract summary, a price change, a post.
• One short brief, written once. The template is below.

Why different companies? Each model was trained differently, so each one misses different things. Asking the same model twice mostly gets you the same answer twice.

The six steps

1. Pick your seats. Two or three apps from different companies. Use the strongest model each app offers you. If it has a "think longer" or reasoning option, turn it on for this.
2. Strip the private stuff. Before anything gets pasted, swap out names, numbers and details you wouldn't hand a stranger. "Client A," "a 12-person HVAC shop," "$X." More on this below.
3. Write one brief. What the work is, who it's for, what a mistake would cost you, and what you want back. Write it once and use it word for word in every app.
4. Paste it to each one separately. A new chat in each app. Don't show one reviewer what another said. If you do, they lean toward agreeing and you lose the point of having three.
5. Ask for a verdict and must-fixes. Every reviewer answers in the same format: SHIP, SHIP AFTER FIXES, or NO, then the fixes. The format is in the brief, so you don't have to ask twice.
6. Compare and decide. Lay the answers side by side. Anything two reviewers flag gets fixed. Anything only one flags, you judge. Then you make the call. The reviewers advise. You decide.

Template: the review brief

Copy this, fill in the brackets, and paste the same filled-in version into each app.

You are reviewing work for a small business before it goes out. Judge it on your own. Don't be polite about problems, and don't approve it just because it looks fine.

WHAT THIS IS: [a reply email to a customer / a price change / a contract summary / a public post]
WHO IT'S FOR: [e.g. a long-time customer who is unhappy about a late delivery]
WHAT I WANT IT TO DO: [e.g. keep the customer, offer one fix, no discount]
WHAT A MISTAKE COSTS: [e.g. losing a $2,000-a-month account]
FACTS YOU CAN TRUST: [the facts the work depends on, in short lines]

Check for:
- anything wrong, unclear, or that could be read the wrong way
- promises I can't keep or claims I can't back up
- tone problems
- anything missing that the reader will ask about

THE WORK:
[paste the work here, with private details replaced]

Answer in the format below and nothing else.

Template: the verdict format

Paste this at the bottom of the brief, right after "Answer in the format below."

VERDICT: SHIP / SHIP AFTER FIXES / NO

MUST-FIX (it doesn't go out until these are done):
1. [the problem] -> [the exact change]
2.

SHOULD-FIX (better, not required):
1.

WHAT I CHECKED AND FOUND FINE:
-

ONE THING I'M NOT SURE ABOUT:
-

The last two lines matter more than they look. "What I checked" tells you whether the reviewer actually read the whole thing. "Not sure about" is where the real problems tend to hide.

Reading the answers

Write down what you decided and why when the reviewers split. After a few weeks you'll see which app catches what, and you can lean on it for that kind of work.

When it's worth the extra step

• Contracts, agreements and anything with a signature on it. The board is a second set of eyes, not a lawyer. If it really matters, a lawyer still reads it.
• Big emails: the unhappy customer, the bid, the "we're raising prices" note.
• Pricing changes and quotes.
• Anything public: website copy, ads, posts, press.
• Any decision you can't easily undo.

Skip it for routine stuff. A two-line reply confirming a meeting doesn't need three opinions.

What not to paste

When you paste something into a chat app, it goes to that company. Treat each one like an outside consultant you haven't signed an agreement with. Leave out:

• Customer and client names, emails, phone numbers and addresses.
• Anything a client sent you in confidence, or anything under an NDA.
• Passwords, account numbers, card numbers, Social Security numbers.
• Health, legal or financial records about a person.
• Employee records and pay.

Swap them for stand-ins ("Client A," "a regional supplier," "$X per month") before you paste. The review works just as well. Also look at each app's data settings to see whether your chats can be used for training, and turn that off if you can.

We learned this one the hard way. This month, a brief that went to two of our outside reviewers quoted a client's own answers word for word. Our client agreements only cover the AI provider they signed up for. So we fixed it in code: now a check reads every brief before it goes out and refuses to send it if a client is named. If you're doing this by hand, the check is you. Read the brief one more time before you hit enter.

Mistakes to skip

• Showing one reviewer another's answer first. Run the first round blind. If you want a second round, share the answers after everyone has given their own.
• Changing the brief between apps. Same brief, word for word, or you can't compare the answers.
• Asking "is this good?" You'll get a yes. Ask for a verdict and must-fixes.
• Letting the board decide. It gives you a better first draft of the decision. You still make it.

This is the same review we run inside CommandKey before client emails, security changes, pricing and launches go out. The difference is that my agent does the pasting, comparing and fixing, and only brings me the decision.

Sources

• CommandKey AI: How we built an AI board that checks our biggest calls
• CommandKey AI: Security