All Things AI · Season 3 · What You Can Build · Episode 6
One chat window is one worker. You can run a group instead, each with its own instructions and its own role, handing work to each other on the same job. It behaves less like software and more like staff, and it is the biggest jump available right now.
Why one model doing everything produces mediocre work
Ask a single model to write something, check it, and improve it, and you will get a polite version of the same output every time. It agrees with itself. It has no reason not to.
This is not a flaw you can prompt your way out of. The model that wrote the draft is the least suited thing in the world to judging the draft, for exactly the reason a writer is a poor proofreader of their own work: the mistakes are invisible from inside the decisions that produced them.
The fix is structural rather than clever. Use more than one, give them different jobs, and make them adversarial where it counts.
What a department actually looks like
The roles are the same ones a real team has, which is not a coincidence. The division of labour exists because it works.
Someone who plans
Takes the request and decides what actually needs to happen and in what order. This is the role most people skip, and skipping it is why so much AI output answers a slightly different question from the one asked.
Someone who does the work
Writes the thing, processes the documents, produces the output. This is the only role most agents currently have, which is why the ceiling feels low.
Someone who checks it against a source
Takes the output and verifies every claim against the actual material rather than against its own memory. This role must be given the source and told to be suspicious, and it must not be the same one that produced the work.
Someone who holds the standards
Reads the finished thing against your rules. Not accuracy, which is somebody else’s job, but whether it sounds right, whether it breaks anything you have said you never do, and whether it is actually good.
Why separating them changes the output so much
Because each one arrives without the context that made the mistake seem reasonable.
A checker that did not write the draft has no investment in the draft. It reads the claim, looks at the source, and either finds it or does not. When that role is folded into the same request as the writing, it inherits every assumption the writing made, and it confirms rather than checks.
The effect is large and it is immediate. The first time you run a genuine separate verification pass over something a model produced confidently, it finds things, and that experience permanently changes how much you trust single-model output.
How you actually set one up
Nothing exotic. The mechanism is just several requests in sequence, each with its own instructions, with the output of one becoming the input of the next.
- Write each role a short standing brief. What its job is, what it is not responsible for, and what it should refuse to pass on.
- Make the checking role adversarial in its instructions. Told to find problems, it finds problems. Told to review, it says this looks good.
- Give the checker the source material, not just the output. A checker with no source is guessing in a more confident voice.
- Run the roles in order and keep what each one produced. When something goes wrong you want to see which stage it went wrong at, and that is only visible if the intermediate work was kept.
Where the leverage really is
Not speed. Work happening while you are not there.
A single chat window requires you present for every step, because you are the thing connecting one step to the next. That makes you the bottleneck at exactly the point where the tool was supposed to remove one.
A group with defined roles and handoffs runs without you between the steps. That is the closest thing to hiring that does not involve hiring, and the difference between the two experiences is not incremental.
The rule that keeps it honest
The checker never fixes. It reports.
The moment a verification role is allowed to correct what it found, it starts making judgment calls about your work without you seeing them, and quiet corrections are how a system drifts away from what you actually wanted while continuing to report that everything passed.
Separate the two. Finding is one job and deciding what to do about it is a different one, and the second belongs to you or to a role you have explicitly given it to. A checker that hands you a list is enormously more useful than one that hands you a silently altered document.
Where this goes wrong
Three failures, and the first two are the most common by a wide margin.
Too many roles. Four is plenty and three is usually right. Every additional role adds a handoff, and every handoff is a place where something gets lost or distorted. A department of eleven produces worse output than a department of three, and much more slowly.
Roles that are actually the same role. If two of them have overlapping instructions they will produce overlapping work and disagree about the overlap, and you will spend your time adjudicating a dispute between two things you built. Each role needs a boundary it does not cross.
Letting one of them be the boss. It is tempting to have a role that decides which output wins. That role is you. Handing final judgment to a model means the thing that goes out has never been read by anyone who is accountable for it.
What it grows
Capacity, in the specific and useful sense of work getting done while you are somewhere else.
Every other capability makes an hour of your time produce more. This is the only one that produces output without an hour of your time in it at all, which is a different kind of gain and compounds differently.
It also raises the floor on quality rather than the ceiling. A single model has good days and mediocre days depending on how the request was phrased. A structured process with a real verification stage produces a narrower range, and the narrow range sits higher than the average did.
What it costs
Four roles is roughly four times the processing of one, which on a local model is free and on a paid service is still small relative to what it replaces. The genuine cost is the hour spent writing the four briefs properly, and those briefs then work for every job you run through them afterwards.
There is a second cost worth naming, which is that a structured process is slower end to end than a single request. Four stages take longer than one. That trade is almost always worth making on anything that goes out under your name, and almost never worth making on something you are going to read and discard in a minute. Use the department for work that matters and a single request for everything else.
Questions agents actually type
Can I use multiple AI models together on one task?
Yes, and it produces substantially better results than one model doing everything. Each gets its own role and standing instructions, and the output of one becomes the input of the next. Planning, producing, verifying, and standards-checking as four separate roles is the common shape.
Why is a separate AI checker better than asking one model to check its own work?
Because a model that produced the draft inherits every assumption that made the mistake seem reasonable, so it confirms rather than checks. A separate role arrives with no investment in the draft, is given the source material, and is instructed to find problems.
How many AI roles should I run?
Three or four. Every extra role adds a handoff where information gets lost, and departments of eight or more produce worse output more slowly than departments of three. Each role also needs a boundary it does not cross, or two of them will do the same work and disagree.
Should the AI checker fix what it finds?
No. Finding problems and deciding what to do about them are different jobs, and the second one belongs to you. A checker that silently corrects things is how a system drifts away from what you wanted while continuing to report that everything passed.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.