How I Write With AI

Most of my writing with AI does not start with a polished prompt.

It usually starts as a long string of thoughts. Half sentences. Fragments. Things I know I want to say but have not organized yet. Sometimes it is dictated. Sometimes it is typed quickly. Sometimes it is just me dumping the shape of an idea before I lose it.

That is the part I do not want AI to replace.

The raw material has to be mine. The opinion has to be mine. The judgment has to be mine. If I do not know what I think, the model can still produce clean paragraphs, but the result usually feels hollow. It might read fine, but it will not be useful.

The useful workflow is different.

I give AI the messy version. Then I use it to find the shape.

The raw material has to be mine

When I am writing a post like this, I am not asking Codex to invent an opinion for me.

I am usually giving it the real inputs:

• What I have been noticing
• What I am unsure about
• What feels true but not fully articulated yet
• What I want to avoid saying because it would be too generic
• What previous writing or project context matters

That last part is important. Good AI writing is not just about the latest model. It is also about the context around the model.

At this point, my agent has a decent understanding of how I work. It has context about my voice, my preferences, my projects, my engineering standards, and the way I want public writing to sound. That context lives in instruction files, repo docs, and durable memory. I am not treating those files as public content, but they matter a lot to the workflow.

That is why the first draft can get much closer than a cold chat window.

What AI actually does

The best use of AI in my writing process is not “write this for me.”

It is more like:

“Here is a messy pile of thoughts. Find the main point. Cut the sludge. Give it structure. Make it sound like me, but cleaner.”

Codex is good at that. It can usually identify the center of the idea faster than I can when I am too close to it. It can see that one paragraph is really the opening, another is the closing, and three other points are probably the same thought repeated in different words.

That is useful.

It saves me from staring at a blank page. It also saves me from manually untangling my own rough notes every time I want to write something.

But I still do not think of that as outsourcing the writing. I think of it as using an editor that is already briefed on how I think.

The context layer matters

This is the part I think people underweight.

A lot of AI writing advice focuses on prompts. Write the perfect prompt. Add examples. Tell it the tone. Tell it the audience. Tell it what not to do.

That helps, but it is still shallow if every conversation starts from zero.

The bigger unlock is durable context. The model should know the constraints before I ask the question. It should know that I do not want corporate filler. It should know I prefer direct, plain language. It should know when a post should be skeptical instead of hype-driven. It should know which projects are active and which details are too private or too noisy to include.

That is the difference between an assistant that can write clean sentences and an assistant that can help me write something I would actually publish.

This is also why my switch from Claude Code to Codex was not only about model quality. I wrote more about that in Why I Switched from Claude Code to Codex. The model matters. A lot. But the context migration mattered too.

If you take away the context layer, the tool gets worse immediately.

Model or memory?

I honestly do not know how much of the improvement is the model and how much is the memory.

Codex has been giving me stronger first drafts than I was getting before. It needs fewer passes. It catches the shape faster. It more often lands in the zone of “this sounds like me after a little cleanup” instead of “this sounds like an assistant pretending to be me.”

Some of that is probably model quality.

Some of that is probably because the context is better now.

And some of it is probably the combination. A stronger model with better instructions and better project memory is going to feel very different from a strong model dropped into a cold session.

That is hard to benchmark cleanly, and I am not sure it matters for my day-to-day workflow. The practical question is simpler:

Does this help me turn real thoughts into publishable writing faster without losing my voice?

Right now, yes.

I still do the final pass

The final pass is where the writing becomes mine again.

I cut lines that feel too polished. I add details that only I would know. I remove anything that sounds like generic AI commentary. I check whether the point is actually true or just convenient. I decide whether the post is worth publishing at all.

That is the part I would not want to automate away.

AI can organize a thought. It can tighten a sentence. It can help me find the opening. It can make a rough idea readable.

But it cannot decide what I believe.

It cannot know which point I am willing to stand behind publicly. It cannot feel when a line is technically fine but not really mine. It cannot replace taste, judgment, or accountability.

That is why I think the best AI writing workflow is still very human.

The point is not to press a button

When I rebuilt my resume, LinkedIn, and website with Claude, the useful part was not that Claude magically knew my career.

The useful part was that I gave it real material. Career notes, project history, old drafts, constraints, taste, and feedback. Then it helped me organize and refine all of that into something usable.

This is the same pattern.

AI is best when it is working from real context. It is weakest when it is asked to manufacture substance from nothing.

So my writing workflow is pretty simple:

1. Dump the messy version.
2. Let AI find the structure.
3. Rewrite, cut, and refine until it is actually mine.

That is the whole trick.

The thinking has to come first. AI just helps me get it out of my head and onto the page without losing the thread.