Writing Might be the Most Important Skill for AI Engineering

Like me, you’ve probably seen too much of Sam Altman. Altman’s company OpenAI and their consumer product ChatGPT are discussed practically everywhere. Even still, I was surprised to discover his recent appearance on the rather esoteric YouTube series How I Write discussing, of all things, paper and pen being his favored tools for thinking. In the video he states:

“For me writing is a tool for thinking most importantly. And I don’t think that is going anywhere, so I think it’s really important that people still learn to write for this reason, in the same way that even if there’s going to be less traditional coding jobs, coding is a great way to learn too.” [1]

I find this a bit shocking — that as more people use ChatGPT in an attempt to enhance their thinking abilities — Altman works through his ideas using an old school notepad. Yet, this notion that writing is a tool to help us think is not novel, associated with idioms such as “white boarding” or “back of the napkin” . Nonetheless, I was struck by Altman’s implication that thinking with writing is essential to our AI future.

It is no accident that OpenAI’s interface is textual, and that we are encouraged to craft our words (or prompt craft) in order to elucidate desired responses from LLMs. If I understand Altman correctly, he believes that writing will remain the interface of choice for generative AI because of how it trains our human minds to produce thought. In other words, while we could use speech or even brain-to-computer interfaces for generative AI, there’s something unique and necessary about how we think when we write.

Articulation

Articulation, in the context of semiotics, or communicating meaning, is the process of joining together the significant elements of a message, combining them to create another layer of meaning. As an example, consider Ernest Hemmingway’s famous six-word short story: For Sale: Baby Shoes Never Worn The elements of the story [“For Sale:”, “Baby Shoes”, “Never Worn”] each have significance, but when combined, create a complex story of human hope and loss. Hemingway’s clever articulation of the elements is masterful.

I’m only touching the surface on the topic of articulation in semiotics⁠ [2], but for our purposes, we just need to understand that units of meaning are combined together to produce new and more complex thoughts.

Consider⁠ the classic meme of a conspiracy theorist (see above image) where newspaper articles and thoughts written down are pinned to a board, labeled and connected by lines of string to each other, thus articulating their overall (and oftentimes complex) conspiracy theory.[3] Similar to Hemingway’s story, the complexity is achievable through the joining of the concepts.

Thinking on Paper

In their book Thinking on Paper Harvard professors Howard and Barton discuss the importance of writing, not for an audience of readers, but instead as a method of thinking through an idea. They point out that “[a] critical part of writing is thinking on paper, or articulating ideas…we do not so much send our thoughts in pursuit of words as use words to pursue our thoughts.”[⁠4] Thinking on paper feels a bit like freeform sketching, or improvising a musical melody, we are not sure what the final product will be, but keep revising until we discover that we have created what we were looking for.

This process of thinking on paper is not just used for articulation, but to actually formulate new concepts. When we combine ideas together, we create new ones that had not been thought of before. We all do this — and it is a bit obvious. But philosophically speaking, it is fascinating that we can teach ourselves new things and alter our minds without any outside input.

It is this learning through articulation, that Altman is referencing when he speaks of his use of a paper notebook. It is more apparent in his later declaration that humans should continue to learn code, even after AI has surpassed our capabilities. The reason is simple, our ability to interact with AI relies on our ability to conceive what we want and communicate it. If we do not learn how to code, we will not comprehend its elements, and consequentially be incapable of the articulation necessary for complex ideas using it.

Prompt Crafting is Thinking

This is where all those prompt crafting and vibe coding tips are leading us. AI’s ability to comprehend and perform tasks that we ask of it is rapidly improving, but our ability to think of more complex thoughts, to think through what we want is stagnant. AI is limited by our own ability to think. However, the feedback loop of writing a prompt and getting back AI generated content is training us to spend time thinking through our thoughts by writing, and to actually structure our sentences differently.

Let’s look at this prompt from How to Create High Quality AI Images Tips for Prompt Crafting:

A playful dark gray cat with soft, fluffy fur frolics in a green park under the warm daylight. The cat chases a fluttering butterfly near a patch of wildflowers. Sunlight filters gently through the trees, casting dappled shadows on the ground. The mood is cheerful and lively, with a realistic, hyperreal, lifelike style, and long-shot view.⁠ [5]

Press enter or click to view image in full size

This prompt provides the AI with units of instruction that the model will use to generate the image [“playful dark gray cat with soft, fluffy fur”, “sunlight filters gently”, “realistic, hyperreal”]. However, the process of writing this prompt required the author to think through their initial idea, spend time conceiving each element, and then articulate them into the overall concept.

As an experiment, verbally describe the image you see. Now look at the image and notice how much less detailed your description was from the prompt above. It would be challenging to produce similar text simply by speaking it. It would be difficult for most to write it on a first pass. It is only through iteration of the prompt that we get one that produces AI generated content the way we wish. In this manner, when we prompt craft, we end up taking time engaging in a type of thinking through writing that is uncommon in most of our daily lives.

Similarly, vibe coding (using AI agents to do the majority of software development) tutorials recommend breaking tasks into small steps, engaging the AI in questions and answers during the process⁠ [6]. Some tutorials even recommend learning the basics of programming⁠ [7], or asking AI to be a project manager and list the small steps it should take. These tips are not just for the benefit of the AI (although they do help), but for the human to better articulate what they are really asking for and how it should work. AI software will improve at producing complex code structures, however humans need to have produce complex thoughts to guide the AI.

Conclusion

If we agree that writing is crucial to more fruitful interactions with AI, especially when it comes to having our true intentions followed, then we can consider methods of improving our writing to be methods essential to AI engineers. In a future post, I will discuss the concepts clarity and meaning in writing software, referencing best practices of the past 80 years that are pertinent for working with AI.

In a time that many, myself included, are concerned about AI produced atrophy of our cognitive abilities, I find the increase in writing to be refreshing. Who knows, maybe it will lead to a world humans spend their time in quiet reflection, reading philosophy and writing in little notebooks to create the world they can dream of.

1 Sam Altman’s Method for Clear Thinking. Directed by David Perell, 2025. YouTube, https://youtu.be/tDmjz6HB-yw?si=Vl9vwH88hAX0lV78&t=124

2 Semiotics for Beginners: Articulation. http://visual-memory.co.uk/daniel/Documents/S4B/sem08a.html. Accessed 26 Mar. 2025.

3 “Pepe Silvia.” Know Your Meme, 24 Mar. 2017, https://knowyourmeme.com/memes/pepe-silvia.

4 Howard, V. a, and J. H. Barton. Thinking on Paper. 1st edition, William Morrow Paperbacks, 1988. p24

5 Malik, Ray. “How to Create High-Quality AI Images: Tips for Prompt Crafting.” How-To Geek, 31 Dec. 2024, https://www.howtogeek.com/how-to-create-high-quality-ai-images-tips-for-prompt-crafting/.

6 Yang, Peter. 12 Rules to Vibe Code Without Frustration. 11 Dec. 2024, https://creatoreconomy.so/p/12-rules-to-vibe-code-without-frustration.

7 Embarrassed_Turn_284. “5 Principles of Vibe Coding. Stop Complicating It!” R/ClaudeAI, 24 Mar. 2025, www.reddit.com/r/ClaudeAI/comments/1jiu7xt/5_principles_of_vibe_coding_stop_complicating_it/.