How I Use LLMs to Write

You've probably seen a lot of skepticism of large language models (LLMs) like ChatGPT and Claude on social media. Often, AI is seen as worthless, as a threat, as a plagiarism machine, as full of hallucinations and so on.

I personally get a lot of value out of using LLMs. I’ve developed a lot of techniques and “mental models” that guide how I use LLMs, from collaborative thinking to avoiding "ChatGPT smell" in my work. I imagine a lot of power users have discovered similar approaches, but since I haven’t seen them written down anywhere, I wanted to share my approach.

A common theme I've seen that unites a lot of fear of AI and a lot of criticism of LLMs in particular is the perception that AI is primarily used to do work “for you”.

I think LLMs require a perspective shift. They won't do a great job doing your work for you, but are an incredible way to help think through the work, create drafts, and help you move iteratively towards a goal.

This distinction should inform how you use these tools. When you treat an LLM as a replacement, you'll get generic, lifeless content with that unmistakable "ChatGPT smell." When you treat an LLM as a thinking partner, you'll maintain agency while exploring ideas more fluidly than you could alone.

Share the work with Shared Context Building

Probably my most important tool is "shared context building." Instead of directly requesting the output I want (write code, an email, etc.) in a single prompt, I ask the LLM to collaborate with me on the output. The goal is to "think out loud" with the model.

I'll typically approach it like: "I want to create [topic or goal]. Ask me questions, make suggestions, and let's collaborate on this."

This approach offers several advantages:

First, having the LLM drive make it easier to discover your own thoughts and goals on the subject. We humans are social creatures; we often think best when another mind is drawing out our thinking. Reacting to the LLM rather than driving the conversation makes it easier to get started, build momentum, and to consider different perspectives on the thoughts.

Second, questions and suggestions help you clarify your own thinking. Explaining your thoughts to an intelligent system forces you to articulate what you really want. And the LLM can often make suggestions that help improve your thoughts by bringing in new perspectives from areas you may be less familiar with.

Third, having the model ask questions has an important practical reason that goes to how LLMs work. LLMs have very strong training to create artifacts early. This is understandable, but users should push back against it and try to defer actual work on the task. A key problem is that a too early draft is too far from the desired work product and it is challenging to review.

When the LLM tries to rush toward creating something too early (they almost always do), it’s useful to gently redirect it back to exploration. I will often just politely acknowledge early work product and steer things back to questions and suggestions.

Sometimes the questions the model asks aren't steering things in a direction that you want to go. That's a sign to understand that the model doesn't understand what you really want, so it's also worth responding to "off" questions by re-affirming high level goals. If the LLM doesn’t understand what you really want, then you probably haven’t found the core thoughts that tie your work together.

When the LLM makes suggestions I like, I’ll often endorse them. An endorsement is generally just a simple statement to the LLM that you like an idea: “I really like the idea about [x].” Later, when creating a draft, I ask the LLM to create a draft that uses my ideas plus ideas that I’ve endorsed. This prevents the LLM from hijacking the draft with ideas it likes but that you didn’t vibe with.

Finding Flow is the Real Goal

The longer I've worked with LLMs the more I've realized that the key goal is to use the LLM to get you into a flow state. That might be different for everyone, but if you don't find yourself getting into that flow state, then it is worth trying something else.

The “ask me a question” approach I described is a great way to find flow. A powerful complementary technique is to start working with the LLM early, before you even know what you really want. When I feel like I can’t find flow or am procrastinating, I’ll often just reflexively open a chat window and start dumping whatever is in my head. It doesn't have to be in any order or have any cohesion; I will start with whatever I have, a topic, a goal, or in the case of this article, I started with some posts I made on Bluesky:

Copycopy code to clipboard
I want to write about how I use LLMs (I apologize in advance that this might get a bit meta.) Here’s some tweets I’ve made:

I’ve been thinking about the same. I imagine power users have rediscovered a lot of the same techniques but I haven’t necessarily seen them written down anywhere.

Probably my most important tool is “shared context building”. Instead of doing the thing I want to do (write code, an email, etc.) I work an artifact towards the thing—a memo, a plan, etc. “I want to create a memo about [x]. Ask me questions, make suggestions, and let’s collaborate on this.

A really great power tool is using text-to-speech with an LLM. Open a window and just start talking. Start working with the LLM as early in the process as you can. It’s a great way to stop procrastinating
——

Here’s how we can work together, you ask me questions, make suggestions, and we will collaborate on this.

You can read the full conversation with Claude that I used to make the first draft of this post here.

This works especially well if you use a text-to-speech tool (I use SuperWhisper on Mac based on OpenAI's excellent Whisper model). Press a button and just start talking with whatever comes to mind. It usually takes me a few minutes to work towards anything useful, but the LLM doesn’t care. Before you know it you’re in the flow.

Starting 'early' often reveals that your initial goal isn't actually where your best work wants to go once you're in the flow. And being in flow is another reason to delay drafts. Drafts require you to go into editing mode, which is always necessary eventually, but interrupts that flow mode. Having the LLM stay in exploration mode with you helps you find what’s really important and key to the work you are trying to do. The question then becomes: how do you know when it's time to move from exploration to drafting?

Creating a Draft

Eventually, you will want to move from exploration to actual artifact creation. I usually know it’s time to start making drafts when I feel like the questions aren't steering things in ways I want to go and I've run out of things I feel compelled to say. If I’ve been in the flow for a while, I’ve usually started to develop a feel for what I’m looking for and have touched on all of those key points in the exploration process.

A key challenge when creating the artifact you want is maintaining your voice. You've probably noticed that LLMs often have a generic way they write that many can now 'smell'. That smell is often a bit annoying; it’s formulaic, cautious, a bit pedantic, in a word, boring. It can also include a lot of structure you don’t like, like excessive section headers and overuse of bullet points and lists.

To avoid ChatGPT smell, I generally ask the model to explicitly use my voice and words: “Please use my voice. Whenever possible try to include my text verbatim (but not in quotes, of course). Where I’ve endorsed your ideas, please include them.”

This instruction will often result in a much more spare draft that only has the bare minimum. But, that draft will be in your words and will be very amenable to further drafts that give the LLM more room to provide input without straying too far from your voice.

Write Your Own Conclusion

I am convinced that you can get benefits from using LLMs if you take the time to understand how LLMs can be complementary to your work. I have presented what works best for me, but the inherent flexibility of LLMs means that you can develop an approach that works well for you. The key is to understand that there is value to be found, and that it is worth investing to find the mental tools that work best for you. Don’t be afraid to write your own conclusion, and have an LLM help you with the rest.