There’s a moment most people have early in working with AI seriously — not the failure of a party trick, but the failure of something that actually mattered. Something you needed. Something you’d invested real time in setting up, and that still came back wrong in a way you couldn’t easily fix.
Mine came about six weeks in.
I was working on one of my YouTube video scripts — something I’d been developing for weeks, with a clear argument, specific examples, a structure I’d worked out carefully. I had a draft I wasn’t happy with and I handed it to an AI to help me get unstuck. I explained what I was trying to do from a structural perspective, pasted the draft, and asked for a revised version.
What came back was competent, clean, and completely wrong. Wrong in a way that was hard to articulate at first, because every sentence was well written and defensible. The argument was clear. The structure was logical. But it had lost the specific thing I was trying to say — the particular angle, the distinctive claim. It had been replaced with something more general, more polished, less useful, and not mine.
What I’d done wrong
When I evaluated my interaction with the AI from an Agile frame of reference, I understood where I’d gone wrong: I’d skipped what Agile culture puts first — setting proper context. I’d told the AI what I wanted done (revise this), but not what the thing was actually for. I hadn’t said: the core argument I’m trying to make is this specific claim. I hadn’t said: the reason the current draft isn’t working is this particular thing. I hadn’t said: here’s what I’ve tried that didn’t work, here’s what I’m preserving no matter what.
I’d handed over the output without handing over the thinking, the desired outcome, the context where this work belongs. And without the thinking, the AI couldn’t preserve what was essential, because I had never said what was essential. The AI optimized for what a readable script looks like in general. That was the only target I’d set.
The lesson I drew
The failure was informative. What it told me about this kind of collaboration was accurate, and not what I’d assumed.
I’d assumed the AI would work like a skilled editor — someone who reads your draft, infers what you’re going for, and helps you get there. That’s not quite right. It’s more like working with a very capable person who only knows what you’ve told them. If you haven’t told them what matters most, they’ll make a reasonable assumption — and reasonable is often not what you needed.
The discipline that follows is front-loading the thinking. Before you ask for help with a draft, write down: what’s the core claim, what’s the specific thing I’m trying to preserve, what would a wrong version look like and why. Not because the AI requires a form to be filled in, but because that thinking is what makes the collaboration useful — and without it, you’re handing over a draft and hoping.
What I do differently now
I’m more deliberate about what I share before I share the draft. The context, the goal, the constraints, the specific things I know are working and the specific thing that isn’t — these go first. That is the other half of what Agile asks for up front: not just the context, but a clear desired outcome. The context says where the work came from; the desired outcome says where it has to end up, and it has to stay visible the whole way through — otherwise the work drifts back toward whatever looks generally reasonable.
Sometimes writing that out reveals that the problem isn’t the draft at all, but the thinking behind it.
When that happens, the AI’s most useful function isn’t editing. It’s being a place to think things through before the writing happens. The failure helped me see that too.
Related content
- Creating Powerful Desired Outcomes: Your North Star for Success — Agile4AI on YouTube

Comments 0