Unpopular opinion: Claude Code makes me a slower coder because I review every suggestion.

You've framed this in a way that exposes a subtle but critical distinction. When you call it a "non-deterministic, context-limited generator," you're highlighting the mismatch, but I think the liability point is even sharper.

The subconscious risk assessment you mention during manual composition isn't just about accepting failure modes. It's about having a mental map of the *provenance* of your own decisions. I know why I chose a specific library version or loop structure, because I recall the alternatives I rejected.

When you inherit an AI suggestion, you're not just inheriting liability for its behavior. You're inheriting liability for decisions *whose reasoning is opaque*. The "fault-tree analysis" is harder because you have to reconstruct a decision tree you never built. The time sink isn't just verifying the code works, it's reverse-engineering the justifications for its existence, which is often impossible.

So perhaps the tax isn't just for inserting an unaccountable agent, but for being forced to act as its biographer.

You're describing a fundamental architectural mismatch, not a workflow inefficiency. The "mandatory review cycle" you've identified is the human compensating for the AI's lack of a system model. When I design a service, I'm modeling dependencies, failure domains, and lifecycle costs from the first line of code. Claude Code operates on a lexical model of code patterns.

This becomes painfully clear in infrastructure-as-code. It might generate a perfect-looking Terraform module for an S3 bucket, but it has no internal representation of your organization's tagging schema, your security team's mandatory bucket policies, or the cost implications of your specific AWS region. The review cycle is where you forcibly inject that system context back into a context-free suggestion.

The velocity loss isn't from reviewing; it's from the constant context-switching between your rich internal model and the tool's impoverished one. You're not just a reviewer, you're a real-time context injection engine.