Writing with LLMs

Feb 2, 2025

It’s not often that we get to experiment with entirely new ways of working. Writing with LLMs is one of those ways. After a few months of using them to write content, here’s what I’ve learned about making them part of my process.

1. Getting Ideas

The ideation phase remains very personal. While LLMs can generate content ideas, I have not felt the need to do that and I would also feel (even) less authentic if I did. Though professional content creators with high-volume demands might benefit from AI-assisted ideation, I find that my challenge isn’t a lack of ideas but rather finding time to develop them all.

2. Exploring an Idea

With an idea in mind, I start with the title and just get some words on the page - LLM-free. Knowing that the text will go through several iterations with later, I feel more freedom to explore various approaches and perspectives. This phase typically produces about 70% of the post’s content. Compared to my pre-LLM writing process, I worry less about perfection in early drafts and allow myself greater flexibility to explore tangents and revise my thinking as the piece develops.

3. Shaping an Idea

After creating a rough draft, I conduct several editorial passes to identify main themes and ensure the narrative holds value. This is where LLM collaboration begins. I share my draft with the AI and pose specific critique-focused questions. For example, rather than asking for general feedback, I might prompt: “Analyze how Section 2 would change if we approached it from a technical rather than conceptual perspective.” This targeted approach generates new insights and improvement areas. To maintain a productive dialogue with the LLM, I frame questions to encourage detailed analysis while acknowledging the AI’s tendency to be agreeable. By requesting specific comparisons or alternative approaches, I attempt to create a constructive feedback loop - not every interaction works, but the overall process always provides more value and leads to a better piece.

4. Refining the Draft

The final stage involves comprehensive stylistic refinement. I present the complete draft to the LLM (particularly Claude) for writing style analysis. Claude is great at identifying opportunities for clearer expression, stronger transitions, and more engaging writing while preserving the original ideas and voice. This stage focuses on enhancement rather than fundamental changes, ensuring the final piece maintains its authenticity while benefiting from AI-assisted polish.

5. Unrefining the Draft

The only problem with an LLM going through the piece is that it looses some of my voice. Small word changes make the piece feel less like me. So after Claude readies the textbook perfect version, I then go through it and essentially re-introduce my voice. I am sure this injects errors back in and inconsistencies with some vocabulary that I wouldn’t have used. However, the piece needs to feel mine and not Claude’s.

Conclusion

That’s my process! It lets me keep control of my writing while using AI to make it better and I get to something faster. I’ve found it’s all about knowing when to bring in the LLM and when to just let my own ideas flow. This way, the final piece still feels like mine, just a bit more polished.

I believe we’re at an exciting juncture in AI development, where agent systems are becoming increasingly important. Through this blog, I hope to contribute to the ongoing dialogue about how we can develop these technologies responsibly and effectively.