How to Write with AI

December 20, 2025

AI Writing Productivity Creativity Thinking Tools

Note:

This is the second out of a three part series exploring the usage of AI to communicate better.

Part 1 talks about why using your voice instead of typing can enhance authenticity.

This article describes a process for writing with AI while preserving authenticity.

Part 3 is a general guide on writing good technical articles.

The Real Question

It is now assumed that everyone is generating code and writing with AI.

The real question is not whether you will use AI, but how will you use AI to stand out.

This article exists because I was dissatisfied with how most people, including myself at one point, were using AI for writing. The promise was speed. The outcome was volume without ownership. What I want to give you here is a process that lets you write fast without losing authenticity. Not borrowed authenticity. Not AI-polished confidence. Your own.Yes, this article was written with AI assistance. But every idea, every sentence, and every nuance is mine. I could defend all of it without AI standing behind me.

The Confusion Phase

The usual workflow looks like this.

flowchart TD
    A[Need to Write Something] --> B[Prompt ChatGPT]
    B --> C[Verbose Output]
    C --> D[Read it Over]
    D --> E[Something Feels Wrong]
    E --> F[Ask for Revision]
    F --> C
    D --> G{Decision}
    G --> H[Accept as Is]
    G --> I[Abandon Draft]
    G --> J[Publish Out of Fatigue after a Number of Revisions]

Even when you get better at prompting, the core pattern does not change. You might ask for an outline first. You might ask for smaller sections. You might tighten the feedback loop. But the author of the text is still not you. You are reacting to sentences you did not think through. You are selecting from options you did not originate.

This is where I hit the first real realization.

Why Prompt Optimization Fails for Writing

For coding and technical artifacts, this workflow is acceptable. There are only so many correct ways to structure a function or a document. Speed matters more than voice. But prose is different. Writing is not just output. Writing is thinking. If the thinking is outsourced, the writing will always feel hollow, no matter how polished it sounds.

That realization forced me to stop optimizing prompts and start redesigning the process.

The core principle I follow now is simple. I must generate the raw thinking. AI is allowed to help me refine it, challenge it, and organize it. It is not allowed to originate it.

Everything that follows is built around that constraint.

Old vs New Writing Model

flowchart TD
    A[Idea] --> B[Prompt AI]
    B --> C[Generated Text]
    C --> D{Accept or Tweak}
    D -->|Accept| E[Publish with weak ownership]
    D -->|Tweak| B

    A2[Thinking] --> B2[Speak Freely]
    B2 --> C2[Raw Transcript]
    C2 --> D2[AI-Assisted Interview]
    D2 --> D3[Refined Ideas]
    D3 --> D2
    D2 --> E2[Structured Writing]
    E2 --> F2[Publish with Strong Ownership]

    classDef Old fill:#f96,stroke:#333,stroke-width:2px
    classDef New fill:#6f9,stroke:#333,stroke-width:2px
    classDef UsesAI stroke-dasharray: 5 5
    class A,B,C,D,E Old
    class A2,B2,C2,D2,D3,E2,F2 New
    class B,C,D,D2,E2 UsesAI

Step 1: Externalize Thinking

Speak, Do Not Type

I use a text to speech tool. It can be ChatGPT voice input or any dictation engine. I close my eyes and I talk. I do not filter. I do not structure. I do not worry about grammar. I dump everything I know, everything I think, and everything I want to say onto the page.

There are two reasons this works.

Speed: speaking is faster than typing. Even a slow speaker outpaces a fast typist.
Cognitive load: speaking removes a cognitive bottleneck. Typing while thinking forces your brain to juggle two tasks that do not support each other. When I speak, my mind stays in thinking mode. The output keeps up.

I've written a detailed article on why free speech production is a powerful cognitive tool here. Read that if you want to understand the theory behind this step.

This step is not about quality. It is about completeness. Digressions are welcome. Repetition is fine. Unrelated thoughts are allowed. Editing at this stage is destructive. It is far easier to cut later than to recover thoughts that were never expressed.

I usually speak for ten to thirty minutes. I do not open my eyes until I am done. Closing my eyes matters because it removes the temptation to self edit or to ask AI for help prematurely.

Most people fail here. If you get this step wrong, everything that follows becomes harder.

Step 2: Research and Discovery

Use AI as an Interrogator

Once I have the raw transcript, I move into research and discovery. This is where AI enters the process, but not as a writer.

Theme extraction: I ask ChatGPT to extract the broad themes from what I have said.
Audience alignment: I ask it to identify the implied audience. I am not checking style. I am checking alignment. I want to see whether what I intended to say is actually present in the material.
Blind spot discovery: Then I ask it to interview me. I tell it to ask questions that would expose blind spots, weak assumptions, or missing context for the target reader. This step is uncomfortable, and that is exactly why it works. Answering these questions forces me to think harder about what I actually believe and what the reader actually needs.

This is where the article usually jumps from mediocre to solid. What felt complete after step one almost never is.

I answer the questions. Sometimes there is back and forth. I stop when the questions become obvious or repetitive. At that point, the content is no longer fragile. It can withstand scrutiny.

Step 3: Structure With Intent

Human Decisions Only

Only now do I move to structuring. I now decide

the themes.
the order.
where the hook belongs and what emotional arc the piece should follow.
whether an anecdote is needed at the start or a blunt statement is better.
how it ends.

AI can suggest improvements. It can point out inconsistencies. It can help tighten language. It does not decide any of these things.

At this stage, AI is a refinement engine. An editor. An advisor. Nothing more.

Every sentence in the final document is something I agree with. More importantly, it is something I could defend without AI standing behind me.

Conclusion: Writing With AI vs Letting AI Write

Karpathy recently tweeted this:

I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become…
— Andrej Karpathy (@karpathy) December 26, 2025

The same applies to writing. If the best in the world are feeling left behind, it is a sign that the the tools are still being discovered and the processes are still being refined. It is fully okay to feel lost right now. Taking shortcuts is easy. And this is what everyone is doing without realizing it. But discovering and following the right process is what will set you apart.

This article is a guide to knowing how to use AI. People call LLMs, Gen-AI. And that's where the pitfall is.

AI is an amplification tool.
AI is a summarization tool.
AI is a filtering tool.
But AI is not a generation tool.
AI is anything but a replacement to your own thoughts.
It works best when it helps you refine, challenge, and organize your thoughts.
It should not originate them.

That is the difference between writing with AI and letting AI write for you.

The speed comes as a side effect. The real win is ownership.