For anything complicated, we never take the first draft from one AI. We use two, from different companies, and make them check each other. Claude writes the draft, ChatGPT reviews it, and the review goes back to Claude for a final version. We call it the three pass method, and we use it for our leases, other complex documents, complex code and every article on our content websites.
How the three passes work
- Draft. Claude, from Anthropic, writes the first version. We run it through the API, mostly from Claude Code, but you could do the same in a chat window.
- Review. The draft goes to OpenAI's model, the one behind ChatGPT, with instructions to critique it: what is wrong, what is missing, what is unclear, what could cause a problem.
- Revise. The feedback goes back to Claude, which works through it and produces the final version.
It sounds like extra work, but the software does the passing back and forth. The only thing that changes for us is that the version we read at the end is better than a single draft would have been.
Why a second AI from a different company
Each AI company trains its models slightly differently, on different material and in different ways, so they have different views. Ask one model to check its own work and it tends to agree with itself. The same goes for different models from the same company. They seem to be trained on much the same material, so whether you use Claude's low cost Haiku model or its high cost Fable model, they tend to agree with each other. That is why we use a second company, not just a second model. Ask a model from another company and it notices different things: a clause that reads two ways, a step left out, a sentence that says more than it should.
The aim is not for one AI to be right and the other wrong. It is the same reason you might ask a colleague to read something before it goes out. A second pair of eyes that thinks differently catches what the first one missed. It is not a guarantee, though. Two models can still make the same mistake, or agree on something that is wrong, which is why anything that matters still gets checked by a person.
Leases: from solicitor to about eight hours
This is where it has made the biggest difference to us. We start every new lease from our existing leases, not from a blank page, so the AI is adapting wording we already know and trust rather than inventing it. The three passes produce the draft, and a person reads every one before it goes anywhere. For property matters there is always a human in the loop.
It has also made us look again at the leases we already had. When we put existing leases, drafted by solicitors, through the same process, the AI review found some serious problems in them. A lease is not safe just because a professional wrote it, and a second reading by two AIs that think differently turned out to be worth doing on our old documents as well as our new ones.
A lease now takes us about four hours to prepare, plus about four hours of negotiation with the tenant's solicitors. We no longer use solicitors ourselves, and that saves us around £3,000 on every lease.
That works for us because we have been letting property for many years and know what a lease needs to say. If you do not, the AI is not a substitute for a solicitor. It will produce something that looks convincing whether or not it protects you, and you need to know enough to tell the difference.
Confidential details never leave our system
Leases and property documents are full of names, addresses and figures. Before anything is sent to either AI, our software replaces each confidential detail with a token ID, and the instructions tell both models to keep the tokens as they are. The tokens pass from Claude to ChatGPT and back again, and only when the final version returns does our software put the real details back in. Neither company ever sees who the tenant is. For the rare document that cannot be anonymised, we skip both of them and use our own AI server in the office instead. We explain the token method in more detail in how we use Claude and ChatGPT safely.
Your situation may be slightly different. ask a question below ↓ and our editorial team will reply with our advice.
Complex code
We use the same method for complicated programming work. Claude Code writes the code, OpenAI's model reviews it, and Claude fixes what the review finds. A second model is good at spotting the cases the first one did not think about, such as missing checks, odd inputs and things that will break later. We use it on the harder parts of our CRM and our websites.
Articles on our content websites
Every rewrite and every new article on our network of content websites goes through the same three passes. Here the AIs do almost all the work. I have a quick look over each article, but it is a final check, not the line by line read and edit a lease gets. The difference is the stakes. A weak paragraph in an article can be fixed later. A badly worded clause in a lease can cost money for years.
Trying it yourself
You do not need to write software to use the idea. The manual version takes a few minutes:
- Ask Claude for a first draft, and give it a good example to work from if you have one.
- Paste the draft into ChatGPT and ask it to critique the draft, not rewrite it: what is wrong, missing or unclear.
- Paste that critique back into Claude and ask it to revise the draft.
- Read the result yourself before you use it, and take out names and confidential details before you paste anything in.
Two AIs disagreeing with each other before you read the result is far better than one AI agreeing with itself.
Ask Technology for Small Business a question
Ask our editorial team a question and we will reply with our advice. Tell us as much about your situation as you can: the more detail you give, the more useful our answer can be.
You do not need to use your real name. Please do not include your full address, phone number, email address, or the names of other people. We may edit or remove identifying details for privacy and legal reasons.
Comments are moderated before publication.