Remember Yu-Gi-Oh? The card game, not the anime—though both make you equally geeky or nerdy, whichever applies. Remember Exodia? And there were a whole set of FTK and OTK decks. First Turn Kill. One Turn Kill.

Lately I find myself thinking about single shot AI capabilities.

Most of the time I work in iteration cycles. Though I keep asking myself: what's the most complex app that can be one-shot?

There's a technical answer. Without subagents, it would have to fit in the context window minus the prompt length. Add in chain-of-thought token burn and you're looking at roughly:

Claude Opus: 200k context, 32k max output. A medium-detail prompt (~1-2k tokens) leaves room for 25-30k tokens of code. That's roughly 1,500-2,000 lines of code in one shot—a full React app with 5-10 components, routing, and basic state. Or a backend with ~15-20 endpoints plus auth.

Past that, you hit coherence issues before you hit token limits. The model starts losing track of its own architecture.

However, that estimate assumes a certain level of detail. There's a spectrum between:

"make 10k MRR SaaS" ←――――――――――――→ "Here's a full design doc, go"

The real challenge is this: the question needs to be answered relative to the initial prompt.

So "notes app" as the broadest, most vague prompt. And a full specification as the least vague. What's the most efficient one-shot approach? Assuming no other AI is consulted, no involvement in planning, with a 100% human-written prompt—what's the ideal level of detail? What does one include?

This goes into the deeper question about how models solve freeform or open-ended problems.

Models have defaults and safe choices, like humans do. They'll pick React/Next.js, or Node, or Python. They'll pick the common language used to solve your problem.

They also need meta context. One often-missed opportunity: telling the model how to best communicate with you. Some people greatly benefit from formality. Others want quick and casual. Some want less technical explanations, or no explanation at all.

Models at this point do not have agentic initiative—or what I'd call a Chain of Initiative architecture. At least not in commercial deployments. They won't ask the meta questions beyond the basics. They might not question scope by default.

I do wonder if single-prompt hackathons are a thing yet.

The best performance I've gotten spends time on the why. The real motivation behind the app. Speaking to a model asking for purely technical output and giving only technical requirements is a larger limitation than most people think.

"Structured XML prompting" is a fallacy in a way. The real gains are from coherent expression of ideas, structured in a way that makes sense. Stream of consciousness prompting is gonna necessitate iteration.

So, can Opus one-shot AGI?

I don't know. But I want to try.