~ Artificial Intelligence Collaboration

What I Delegate and What I Don't

People ask me how I use AI. The honest answer is: constantly, but with a very specific division of labor that I have refined over hundreds of projects. The line between what I give to the model and what I keep for myself is not arbitrary. It reflects a theory about what humans are good at, what AI is good at, and where the boundary between tool use and abdication falls.

Here is the concrete accounting.

What AI Gets

First drafts. Almost all of them. I used to stare at blank pages. Now I describe what I want and let Claude generate a starting point. The starting point is never the finished product. It is a block of marble that I carve. But having a block of marble is vastly better than staring at an empty quarry.

Code scaffolding. When I need a Flask route, a database schema, a utility function, I describe what it should do and let the model write the first version. I read every line. I modify most of it. But the typing is not where the value of my time is. The design decisions and the quality review are where my time belongs.

Research synthesis. When I need to understand a topic that spans multiple domains, I ask the model to summarize the landscape. What are the major positions? What are the key papers? Where is the disagreement? This is not a replacement for reading the primary sources. It is a map that tells me which primary sources to read first.

Format conversion. Taking an essay and restructuring it for a different audience. Turning a technical explanation into something accessible. Adjusting tone without changing substance. These are mechanical transformations that the model does well and that I find tedious.

Error checking. Proofreading. Catching logical inconsistencies. Flagging places where an argument is weak or a claim is unsupported. The model is a good first-pass editor. Not a replacement for human editorial judgment, but a filter that catches the obvious problems before a human needs to engage.

What I Keep

The thesis. The central argument of any piece of writing is mine. I decide what I believe. I decide what I want to say. The model can help me say it more clearly, but the decision about what to say is not something I outsource. If I cannot articulate my own thesis without AI help, I do not understand my own thinking well enough to write about it.

Voice. This is the hardest thing to describe and the most important thing to protect. My writing sounds like me. It has a rhythm, a set of preoccupations, a way of approaching problems that reflects decades of thinking and living. The model can approximate my voice if I give it enough examples, but approximation is not the same as authenticity. When I let the model's voice replace mine, the writing becomes competent but hollow. The ideas are present but the person behind them is absent.

Editorial judgment. The model generates options. I choose between them. This sounds simple but it is the highest-leverage activity in the entire process. The difference between a good essay and a mediocre one is rarely in the quality of the sentences. It is in the decisions about what to include, what to cut, what to emphasize, and what to leave implicit. These are judgment calls that depend on understanding the audience, the context, and the purpose in ways that the model cannot fully access.

Emotional truth. When I write about schizoaffective disorder, or about struggling with the ethics of technology, or about what it feels like to build something you care about, the emotional content must be mine. The model can help me find the words. But the feeling that the words point to has to come from lived experience. When I let the model supply the emotion, the result reads like a greeting card. Technically appropriate sentiment without any actual weight behind it.

Ethical positions. When I critique algorithmic optimization or argue for a particular approach to AI development, those positions are mine. I take responsibility for them. Delegating your ethical stance to an AI is not efficiency. It is evasion.

Why the Line Matters

The distinction is not about pride or purity. It is about quality and integrity. The things I delegate are things where the model adds speed without losing substance. The things I keep are things where my absence would be detectable, where the work would be technically competent but missing the thing that makes it worth reading.

I have seen what happens when people delegate everything. The output is fluent, structured, and empty. It has the shape of thought without the substance. You can produce enormous volumes of it and none of it matters. I have also seen what happens when people delegate nothing, insisting on doing every keystroke themselves as a point of principle. They produce less, work harder, and the quality is not meaningfully better because they spent their energy on the parts that did not need their attention.

The goal is to spend human attention where it matters most and use AI for the rest. This is not a formula. The line moves depending on the project, the stakes, and how I am feeling on a given day. But the principle is constant: delegate the labor, keep the judgment.