Slop isn't just a written-word problem — it's a workflow problem too. A Plain English episode on AI detection got me down this road, though I ended up somewhere different than where that conversation went.

I keep building agentic systems, and one-shot prompting keeps getting better — hand a model a goal, let it run, it gets there. That part's real. But the more I build, the more I notice these things pushing in the wrong direction. Three layers deeper into some corner of a problem than any actual domain expert would bother with. Building complexity nobody asked for because it's possible, not because it's right.

Prompt engineering was a skill. Context engineering became one. Harness engineering is turning into one now too. I think curation is next — knowing when to stop, what to leave simple, what actually needs to exist versus what's just showing off. Clean architecture first, the simple version first, growing from there instead of piling on.

Where this probably goes: a small slice of people who get genuinely good at that — curating the minimalist path on purpose, on top of everything these models can now do. And a much bigger slice of bots and agents running wild with basically no curation at all, which is most of the slop already out there. Feels like we're right at the start of that gap opening up.