Whenever two of us at Techstars disagree about something real, we do what we’ve started calling a Claude mind meld. I work the problem in my Claude. The other person works it in theirs. Then we pass prompts and outputs back and forth until the two sessions are arguing about the same thing in the same terms.
It sounds like a gimmick. It isn’t. It has become the fastest way I know to get to the actual disagreement.
Here’s the example I keep coming back to. Our product marketing says the product does a set of things. The codebase says something different. Not wildly different, but different in the places that matter, and both sides are completely sincere. Marketing isn’t lying. Engineering isn’t sandbagging. The docs describe what was true, or what was almost true, or what shipped behind a flag that nobody turned on.
That argument used to take weeks, because nobody can hold both artifacts in their head at once. The marketing pages are long. The repo is longer.
So we mind melded it. One session working from the marketing claims. One session working from the actual code. Then back and forth. Line up every claim against what the code really does, and say plainly which of three buckets it’s in: true today, partly true, or not true.
Most of the gap burned off fast. A lot of it was vocabulary. Marketing calls something an integration and engineering calls it a webhook, and both are right. Some of it was staleness, claims that were true two releases ago. The models are relentless about naming that stuff out loud, which humans in a meeting are not, because naming it feels like an accusation.
What was left was three real disconnects. Three places where the product genuinely does not do the thing we tell people it does, and no amount of analysis resolves it, because the only question left is what we’re going to do about it. Change the page, or build the thing. That’s a business call. That’s mine to make.
Three decisions, clearly stated, in an afternoon.
Then we made the calls. And once we did, both sessions reconciled against each other almost immediately, and each one explained the outcome back to its own human in that human’s frame. The marketing side got it as a copy change and a date. The engineering side got it as two tickets and a de-scope. Same decision, understood on both sides of the table, with nothing lost in translation.
Now the part that matters most. We are not sitting back and letting two bots debate each other. That produces fluent mush, and I’ve seen plenty of it. At every pass, both of us are overlaying our own domain expertise. I’m correcting what my Claude assumed about how we actually sell. The other person is correcting what theirs assumed about how the code actually ships. The models are doing the drafting and the reconciling. The humans are doing the knowing.
From the outside it can look like one Claude talking to another Claude. It isn’t. There’s an intelligent human on both ends at every single step.
This connects to something I keep noticing about the people who are using AI well, which is that they’ve changed the shape of their work.
They spend almost all of their energy on the first 20% of a project and the last 10%, and they hand the middle to the machine.
The first 20% is framing. What are we actually trying to do, what do we already know, which constraints are real and which ones are just habit. That’s thinking, and it doesn’t get easier because you have a model. It gets more important, because a weak frame now produces a confident 80% of work that’s wrong in a way that looks finished.
The last 10% is judgment. Is this right, would I put my name on it. You can’t skip that and you can’t delegate it.
The middle 70% used to be the job. For a lot of work, it isn’t anymore.
The mind meld is that same ratio applied to a disagreement between two people instead of to a project. We do the framing. The machines do the grinding. We make the call at the end.
