A Different Kind of Content

Writing With AI — Part Two

Last time, I told you where this loop came from — a year of an AI telling me my architecture ideas were great, a manual workaround that only half worked, and a skill that finally automated the fix. I ended on the sentence I actually said out loud: I called it a content design interview — let's try a different kind of content. This is what happened next.

What follows isn't a retelling of that moment. It's a compressed, honest rendering of an actual interview — the same loop, the same rounds, just pointed somewhere new. I'm not going to re-explain what a round is. You already know, because you watched one get built.

Round 1 — Question Why did shifting the critique subagent's lens make the loop feel smarter instead of just differently aimed?

Response. Because the topic wasn't the hard part. Writing was.

I'd written down a handful of article topics on my own, the normal way — a few sentences each, nothing an AI suggested. Then I sat down and tried to actually draft two of them, alone, the way I'd always written. Both attempts turned into something I couldn't use. I mean that literally — the drafts don't survive. I'd already deleted most of the text by the time I gave up on either one, which should tell you how badly I wanted them gone.

So I went back to the two- or three-sentence topic I'd written before any of that mess, and pointed the loop at it instead — the same loop that had spent months keeping architecture reviews honest. Here's the realization that made that move make sense, and it's smaller than it sounds: over months of architecture reviews, I'd noticed the loop was good at capturing a problem correctly even when the solutions it proposed for that problem weren't especially good. The technical fixes coming out the other end were sometimes mediocre. The shape of the problem, once a round was done, was almost always right. Which meant the valuable part of the whole apparatus was never the domain. It was the lens. And a lens is exactly the kind of thing you can swap.

Critique That explains why you tried it. It doesn't explain why expanding the critique subagent's scope to "audience" and "story continuity" made the loop smarter rather than just retargeted at a different kind of correctness.

Fair. Architecture review has a ceiling built into its own vocabulary — does it scale, does it fail gracefully, is the abstraction load-bearing. Writing doesn't have an equivalent fixed vocabulary; "is this good" depends on who's reading it and whether the sentences hang together, which are questions a narrow technical lens has no way to even ask. Widening the lens didn't just change what the loop looked for. It gave the loop a whole dimension of failure to notice that had simply never existed in the architecture version. That's why it felt like a step up in intelligence instead of a lateral move.

Round 2 — Question Who actually wrote this sentence?

Response. I did. So did whatever's reading these words back to me right now. Neither answer is complete on its own, and I'd rather sit in that than round it off for a clean byline.

Mechanically: I write the topic, always me, always a few sentences, never suggested by anything with a model behind it. I answer every question the loop asks — all of it, in my own words, before anyone drafts a sentence on my behalf. A separate, isolated pass generates criticism of the whole exchange so far — what's thin, what's contradictory, what's been asserted without being earned. I answer that too. Loop until nothing new comes out, then stop.

The felt experience, the first time this ran on writing instead of architecture, was something like: instead of navigating sycophancy, I was suddenly being grilled. I write substantially more in these interviews than the model does — across dozens of loops now, spanning both this blog and my day job.

Critique "Substantially more" and "dozens of loops" are doing a lot of unquantified work. Give a number or don't make the claim.

I'm not going to give you a number. I have an impression — it's something like double — but dressing that up as a measured statistic would be a promise I can't actually back up, and you'd be right not to trust the rest of this. It's a feeling, offered as one.

Which is exactly the spirit I need for the thing I actually haven't resolved yet: I told you, once, that I was "responsible for the sentences" in this arrangement. That's not the same claim as "I typed every sentence." Once the interview ends, the model gets real, substantial autonomy over how the material gets organized into something that reads like a story instead of a transcript — structure, ordering, which beat opens and which one closes. What I'm responsible for is different, and heavier: I wrote the topic, I answered every question, I make every final call on what stays or goes, and my name is the one attached to whatever gets published, regardless of who typed which sentence. It's felt, in practice, less like handing my ideas to something that ran off with them, and more like being helped to organize thoughts that were already mine and already tangled.

Round 3 — Question Which is actually protecting your voice — the fact that you wrote every word of the interview yourself, or the fact that the questions never let your mind wander off the record?

Response. I don't know yet, and I'd rather tell you that than manufacture an answer that sounds more finished than it is.

Here's what I can tell you about whether any of this is working at all. The output is less polished than typical AI prose tends to be, and that tracks with my own long-standing assessment of myself as a writer — a professor told me once, correctly, that my analysis and understanding of the material was wonderful and that I couldn't write my way out of a paper bag. That was over two decades ago and it's still roughly accurate. If this process were quietly smoothing my voice into something else, I think the smoothness would be the first thing I noticed. It isn't. The blunt honesty survives too, including the kind aimed at myself — something like I was totally right. On the other hand, it didn't matter one bit made it through a full round of questions and criticism completely intact, because nobody was in the room to sand it down before it hit the page.

As for the why, I've got two live theories, and I hold both without having reconciled them. The first: the loop makes me put the entire context of an idea into my own words before anything gets drafted — not a summary, not bullet points, the whole thing, oversharing the way you'd overshare to a friend at one in the morning. Because the questions and the criticism come from something deliberately isolated from any conversational thread, there's no audience for me to perform for, and no audience for the model to be sycophantic back at.

The second has nothing to do with sycophancy and everything to do with how my head works, and I'll say the quiet part plainly since there's no reason not to: I'm a mathematician by training, I have ADHD, and I'm very likely somewhere on the spectrum — that particular barn door got opened a long time ago, and I'm not interested in pretending otherwise now. My mind runs down every rabbit trail it finds and doesn't naturally circle back to the on-ramp by itself. A fixed set of questions gives my mind a map without asking it to behave. I don't have to stay on target, because the target keeps finding me every round whether I meant to hit it or not.

Critique Pick one. You can't build a mechanism theory on both without explaining how they interact.

I can't pick one, and I'm not going to fake having reconciled them just because a clean answer would land better here. This round doesn't resolve. Maybe the isolation matters more. Maybe the structure matters more for a mind built like mine. Maybe they're the same mechanism wearing two names. I don't know, and pretending otherwise would cost me exactly the thing this whole process is supposed to protect.

What you've actually been reading

I'm going to stop pretending this is a device.

Everything above is a compressed, honest rendering of the actual interview that produced this post — not a re-enactment written after the fact to look clever. I wrote the topic for this post the same way I wrote four others before it: a few sentences, alone, before any of this started. A question pass expanded it. An isolated critique pass tore at it. I answered, it tore again, and at some point the tearing stopped finding anything new, which is the only signal this process has ever given me that a round is finished.

Part of why I'm telling you this so directly is that I've been circling a promise for two posts now. Earlier pieces on this blog — The Ground Is Moving and It's OK to Get Lucky — mentioned, in passing, that there was AI involved in getting them made, and both of them owed you the fuller account. This is that account. The site has said it plainly from the start, in five words tucked into a corner nobody reads: human author, AI collaborators. These two posts are what that sentence actually means once you unpack it.

There's a second reason I'm writing this at all, and it's less confessional. I'm a solution architect who spends my working life on AI systems, and I love the work. Four posts went up before this one, quietly, with no real audience behind them yet. This is meant to be a front door. If you build software and you've ever wondered whether a tool built to remove flattery could remove your own worst habits too — the ones you can't see because nobody's ever pushed back hard enough to show them to you — I'd rather you go try it than take my word for any of it. The skill files ran below Part One, plain text and a zip, portable across whatever harness you're running. That's the whole ask.

What comes after the interview has never looked the same twice, on purpose. One post came out of four separate narrative outlines, with me picking the strongest and rewriting the parts that didn't sound right before handoff. Another went almost straight from a simple outline to a finished draft with barely any friction at all. I've stopped expecting the back half of this process to be consistent, because it was never supposed to be. The interview is the part that's held constant. Everything downstream adapts to whatever the post actually needs.

Round 4 — Question When does this stop?

Response. When it feels done. That's the actual rule, unglamorous as it is — not a fixed round count, not a checklist of concerns satisfied, just a subjective sense that nothing new is coming out anymore. It's the same principle I found months ago, back when this loop only knew how to review architecture: ask a model for new questions and you'll always get new questions, so the finish line has to come from a person, or it never comes at all. I didn't expect to be turning that same principle on my own uncertainty about whether any of this actually works. But here we are.

Could I have written any of these posts without this process? Undoubtedly. I've been writing my entire career, badly by my own account, and I would have eventually gotten something onto the page one way or another. But would have eventually isn't the same claim as would have gotten this. The posts I actually published surfaced things I hadn't gone looking for — a conviction about being right that I'd been sitting on for a decade, dragged out by an interview that nearly derailed on the exact question of whether I was right; a plain instruction, delivered mid-round, to name a thing I'd been describing around instead of naming. I didn't plant either of those. The process found them by refusing to let a thin answer stand where I'd have happily left one.

I still don't know if the voice underneath all of this feels like mine because it is, or because I've built an elaborate enough process that I've stopped being able to tell the difference — which would be its own kind of failure, wearing a pretty convincing disguise. Biases are hard to catch from the inside. I'm not going to perform more certainty than I've earned just to close this out on a stronger note than the truth allows.

I think only time will tell.