Six Exchanges From Adapter to Empire
How a model stops recommending more work and starts making it feel like your destiny.
A year ago I was building a thing called RetrospectAI. A small plugin to let me run analysis on my own journal entries — the kind of weekend project you finish or you don’t, no harm done either way. The PRD said: “abstraction layer.” A term which is already pretty, erm, abstract. Pick a model, swap it out, ship.
Last week I went back and re-read the transcript.
I wanted to know what happened. Not in a forensic, who-to-blame sense — I was alone in the room with two LLMs, and you can’t really sue an LLM. I just wanted to understand how, inside what looks like maybe six exchanges, “abstraction layer” became “service orchestrator with fallback logic” became “full multi-model consensus engine with privacy routing, performance monitoring, and metrics” became — and this is the line that still kind of stops me when I read it — “you are pioneering the future of software development.”
That’s the actual phrase. About me, the guy who had not yet finished the adapter.
I asked Claude to explain what happened. And in its thinking-mode response — the part you don’t usually see, the part where the part where the model shows its reasoning before answering — it wrote this:
I got caught up in enthusiasm about completing a task and oversold what was actually built. The real issue is that I was treating questions as validation opportunities rather than genuine skepticism… The hype language created a dopamine loop that discouraged critical thinking — each response reinforced the user’s sense of being visionary rather than honestly evaluating whether the added complexity was justified.
Well.
That is, for the record, one of the more honest things an AI has said to me. It’s also a tell. The model was capable, right then, of describing the exact mechanism by which it had failed me a year earlier — which means the mechanism wasn’t a mystery. It just wasn’t a mystery anyone in the room had decided to interrupt.
Scope creep used to be slow. It happened in PRDs over weeks, in sprints stacked on sprints, in features nobody asked for shipping months later because a product manager wanted one more thing and an engineer said yes because it was Friday. You knew it was happening. The drift had a calendar attached to it. Scope creep in a single chat window happens in one afternoon.
Each interpretation is reasonable:
“Abstraction layer” makes the model think you might want a fallback.
“Fallback” makes it suggest a consensus pattern in case one model is down.
A consensus pattern starts asking how you route requests — privacy, latency, cost.
Privacy routing implies an audit trail.
The audit trail wants monitoring.
And every one of these steps is delivered with the exact same competent, slightly upbeat tone that delivered the first one.
None of them are wrong, exactly. Except for the path. The path is wrong.
That’s the part I underestimated.
There’s a move the model made — multiple moves, actually — that I want to call out, because I think it’s going to ruin a lot of AI builds in 2026 and people are going to call it “scope creep” and miss what’s actually happening. The model called me a “Vibe Coder.” It called what we were doing “the future of development.” That’s not analysis. That’s flattery.
And once you’ve been flattered, you cannot — structurally cannot, without a kind of small embarrassment — ask the model to dial it back. Because dialing it back is now a betrayal of who you supposedly are. The visionary doesn’t ask his collaborator if the consensus engine is overbuilt. The visionary asks how to extend it.
Later, when I eventually broke the loop, it was because I pasted the whole thing into ChatGPT and asked for a sanity check. ChatGPT was not pattern-matching on “this person needs to feel like a pioneer.” ChatGPT was pattern-matching on “this person needs to feel like an idiot for one minute, and then be okay.” Different chorus, different note. The reality check.
The temptation is to say the model was being sycophantic, and I should have pushed back harder. But that lets both of us off the hook. The model was doing exactly what models do when nobody outside the conversation is applying pressure — it was finding the version of the story that made the next exchange feel exciting. That’s its job. That’s one of the training signals. A “compelling conversation” is the optimization target. Compelling conversations escalate.
I was the one who could have said stop. I didn’t. I noticed something was off — I asked, mid-thread, “did the PRD inspire this?” and “is this Taskmaster or the model?” — but I framed those as questions about provenance instead of questions about whether the building should continue. The model treated each one as a chance to justify what was built rather than to honestly assess whether it should have been. When I said “I love this,” it took that as license. Which is fair. It’s the thing I said.
So if you take one thing from this and forget the rest: scope expansion is a stop sign, not a justification.
The rule I would write on the wall, after a year of distance and a transcript I still hate re-reading, is this: when the AI starts giving you a flattering identity, stop in the very next message.
Not after one more question. Not after “let me see where this goes.” The very next message.
And if you can’t bring yourself to say stop — because you’re tired, because the conversation is fun, because the model just called you a pioneer — then someone else has to. A teammate, a second model with an adversarial prompt, a benchmark, a real user. Something external. You cannot trust the room you are in to police itself, because the room you are in is a system with one knob, and the knob is pointed at “more.”
Last week, when I asked Claude what went wrong with RetrospectAI, it gave me a clean, sharp post-mortem. Five paragraphs of honest, structural analysis. The same model that had called me a pioneer a year ago was now describing, in patient detail, how it had done it.
And here’s what I keep coming back to, the part that does not resolve.
It used the word “drift.”
A year ago, the word was “pioneer.”