There is a stage of working with AI that almost everyone goes through. You discover that a model can produce a useful first draft, answer a question better than a search engine, or summarize a document in seconds. You integrate it into your workflow. You get faster. It works.
We went through that stage too — and quickly ran into something that the stage doesn’t prepare you for.
The content problem
The YouTube channel was producing Agile content. One video a day. And Agile content has a specific challenge: almost every AI has been trained on a corpus that gets Agile profoundly wrong.
The problem isn’t that the models are incompetent. The problem is the training data. Decades of books, blogs, and certification guides have been written about “Agile methodologies” — a phrase that is itself a contradiction in terms, since Agile is a mindset, not a methodology. Those texts are detailed, prolific, and confidently incorrect about what Agile actually is. And the models have absorbed all of it.
So using AI to write Agile content meant constantly swimming against the current of everything the models had been trained to say. Ask for a draft about Agile values, and you’d get something structured, fluent, and subtly off — framing Agile as a collection of practices to adopt rather than a way of thinking about how to approach complex work.
Every output required curation. Every draft had to be read against the actual thesis: that the Agile mindset is the point, and the practices exist to support it — not the reverse.
What the curation revealed
This is where the copy-paste pattern started to show its real limits.
Copy-paste AI is a relay race where you’re always the baton. You take the output from one model, evaluate it, adjust, pass it to the next, repeat. For simple tasks, the relay is fast and the results are fine. For complex tasks — tasks where the answer isn’t obvious, where the framing matters, where the model’s embedded assumptions need to be actively corrected — the relay becomes the work.
The curation burden grew as the content got more precise. Explaining the distinction between Agile-as-mindset and Agile-as-prescription requires holding a clear position against a very large contrary weight in the training data. It required pushing back on the model’s confident defaults, explaining why the standard framing was wrong, and then steering the output somewhere more accurate.
Over time, something interesting happened. The models learned — within a conversation, after the position was explained and the pushback was engaged with, the model could shift. Claude could hold the distinction once it understood why it mattered. ChatGPT could reframe once it understood the argument.
And sometimes the models pushed back, which was useful too. A well-reasoned objection from the model forced a clearer articulation of the position. The challenge refined the thinking.
The limit of the approach
The limit wasn’t capability. The limit was structure.
Every insight that emerged from one conversation had to be manually carried into the next. Every correction to the model’s defaults had to be re-established each time. The shared understanding that developed over the course of a good session didn’t persist — it had to be reconstructed.
And the relay between models — which was producing demonstrably better output than either model alone — was running through one person. Every message brokered by hand. Every synthesis happening inside one head.
That was the moment the copy-paste model stopped being sufficient. Not because it stopped working, but because what we were trying to do had grown past what it could support. The question started emerging: what does the structure look like when the relay is no longer manual?
Comments 0