How Math Makes Art: Behind the Scenes of AI-Assisted Writing
A peek inside my two-year experiment with AI-assisted writing and the tools that power my weekly articles
Last week, I explored whether math could make art—examining the philosophical question of whether AI creativity is fundamentally different from human creativity. I argued that what we call "creation" might actually be sophisticated search, whether performed by humans or machines.
But there's a meta-layer to that article I didn't mention: it was itself a collaboration between human and AI. There are many great reasons not to use AI for writing. This isn't about that.
I started this blog about two years ago with two specific goals. First, to learn about AI—publishing weekly gave me a natural deadline to explore everything from copyright implications to the geopolitics of AI innovation. Second, to learn how to use AI—I see it as an evolving lump of clay, and hands-on time with a specific goal is the best way to develop a feel for this new medium.
For two years, I've had my hands in this clay. I've felt it change consistency as models improved. I've developed techniques for molding it more effectively.
In this article, I'm giving you an overview of my process. I'll walk through how last week's piece came together—from initial conversation to published article.
My Tools: The AI Workshop
Here's an overview of the key tools in my AI writing workshop—the components that work together in my creative process.
Core Writing Environment
Cline: A VSCode plugin that serves as my primary AI assistant. At first this might seem like an odd choice—Cline is designed for coding, not writing. But software development workflows come with many features essential for writing: version control, separation of concerns (content vs. styling), and efficient file management. I write everything in markdown, which makes it universally portable, and Cline excels at reading and manipulating these files. What makes it powerful is its ability to load custom system prompts for different projects—my writing assistant uses entirely different instructions than my coding assistant.
BoltAI: My multi-LLM chat interface with built-in Whisper transcription. I use this primarily to capture my initial thoughts through voice, letting me brainstorm out loud and have those ideas automatically transcribed. The multi-LLM capability lets me switch between different models depending on the task. I initially wrote drafts directly in BoltAI, but over time I've migrated most of my writing workflow to Cline for its deeper integration with my file system.
MCP Servers for Extended Capabilities
While Cline and BoltAI form the core of my writing environment, they'd be limited without access to external information. This is where the Model Context Protocol (MCP) becomes essential—it's a framework that lets AI tools connect to external services. I use several MCP servers to extend what my AI assistants can do:
Brave Search: Gives my AI assistants the ability to search the web for current information, research topics, and verify facts. This keeps my writing grounded in accurate, up-to-date information without constantly switching between my editor and browser.
Replicate with Flux 1.1: Enables image generation capabilities directly from my writing environment. Flux 1.1 is a powerful text-to-image model from Black Forest Labs that I use for article illustrations, allowing me to visualize concepts without leaving my workflow.
FireCrawl: A web scraping tool that can fetch entire webpages and convert them to markdown. This lets my AI assistant read and analyze external content without me having to copy-paste, making research seamless and maintaining my creative flow.
Knowledge Management
Beyond the tools that help me write, I need a system to organize what I've learned and the connections between ideas:
Obsidian: My "Second Brain"—a markdown-based note-taking system that creates a network of interconnected notes. I store research, ideas, and reference materials here, which I can easily pull into my writing process. Obsidian helps me build connections between ideas across different articles and projects.
This toolset has evolved over two years of weekly publishing. Each component plays a specific role, but the real power comes from how they work together—creating a system that's greater than the sum of its parts.
Anatomy of an Article: How "Can Math Make Art?" Came Together
Let me walk you through how last week's article came to life, from initial spark to published piece:
It started at a party last Saturday. I found myself in a conversation about AI and creativity. Someone expressed skepticism that AI could truly create anything original—it was just "auto-complete," they said, a pattern repeater incapable of genuine creativity.
The conversation stayed with me. There was something about the distinction between "searching" and "creating" that felt artificial, but I couldn't immediately articulate why. I decided to explore it in my next article.
Sunday Ideation: Capturing Raw Thoughts
The next day, I opened VSCode and created a new folder called "search-or-create" in my repository of Substack articles. I created a markdown file for my initial notes, then used BoltAI's Whisper transcription to capture my stream-of-consciousness thoughts on the topic.
I spoke for a few minutes, exploring the idea that what we call "creation" might actually be a form of sophisticated search—whether performed by humans or AI. The transcription wasn't perfect, but it captured the essence of my thinking:
From Raw Ideas to Structured Outline
Collaborative Outlining
Next, I launched Cline and asked it to help me outline the article. This is where my system of prompts comes into play. When Cline starts up in my writing project, it automatically:
Reads my system prompt that explains my writing process
Loads my style guide that helps eliminate "AI-speak" and reinforces my conversational style
Has access to my previous articles to understand my voice and approach
I worked with Cline to refine the outline, structuring the piece around a central question: Is there a meaningful difference between search and creation? We developed a flow that would start with a thought experiment for the reader, then explore the idea that all creativity might be recombination of existing elements.
A key part of this was refining the three stages of the thought experiment. My initial notes from the transcription were quite raw. Cline helped me structure how these would be presented. For instance, when we were trying to define the third, most 'creative' scenario for the experiment, our conversation went like this:
PS: Conversations have been edited for clarity and brevity.
Cline: For the third example, we could push into what most people would consider "pure creation" territory by asking readers to: "Imagine an alien species that has evolved on a planet with completely different conditions than Earth. Describe what this species looks like, how it communicates, and one unique aspect of its culture." ... What do you think of this example?
Me: I like it, but let's stick with the restaurant theme. Let's say you went to an alien restaurant on a space ship. What would the restaurant look like? What would be on the menu?
Cline: That's perfect! Keeping the restaurant theme creates a stronger narrative thread while still pushing into more creative territory. I really like this progression... This sequence provides a perfect setup for discussing how both human and AI creativity exist on a spectrum of search-based processes.
This kind of back-and-forth was typical. Cline would offer a suggestion based on my initial input, and I would then steer it, often to maintain thematic consistency or to align the idea more closely with my core argument. In this case, Cline's initial idea of an 'alien species' was good for prompting 'creation,' but my refinement to an 'alien restaurant' kept the experiment tightly focused on the food/restaurant motif we'd already established.
Research and Reference Collection
With the outline in place, I needed supporting material. I opened my Obsidian vault and searched for notes related to creativity and remixing. I found references to Kirby Ferguson's "Everything Is A Remix" documentary and the SCAMPER technique for creative thinking.
When I shared these concepts with Cline, it suggested ways to develop them more fully:
Me: (In my notes) Everything Is A Remix is a documentary web series by filmmaker Kirby Ferguson... techniques like SCAMPER: S » ubstitute, C » ombine, A » djust, M » odify/Magnify/Minify, P » ut to Other Uses, E » liminate, R » everse/Rearrange
Cline: The Steve Jobs quote is perfect as an opener. After that, we could:
Expand on Kirby Ferguson's documentary with specific examples he uses (like Led Zeppelin's borrowing from blues musicians or Apple's GUI borrowing from Xerox PARC)
With SCAMPER, we could provide a concrete example of how this works in practice - perhaps taking a familiar concept (like a restaurant) and showing how each SCAMPER technique transforms it into something that feels "creative" but is actually a systematic transformation
For the creative works examples, we could dig deeper into one example (maybe Star Wars) to show specifically how it remixed elements from Flash Gordon, Kurosawa films, and Joseph Campbell's hero's journey
This exchange shows how Cline doesn't just passively receive research material—it actively helps develop and expand on concepts, suggesting specific examples and applications that strengthen the article's argument. The AI suggested concrete examples like Led Zeppelin and Star Wars that I hadn't specifically mentioned but that illustrated the remix concept.
I also used FireCrawl to retrieve information about Jorge Luis Borges' "Library of Babel" concept, which outlined the idea that all possible written works already exist in theory.
Drafting and Refinement
With my outline and research in hand, I worked with Cline to develop each section. The process was collaborative and iterative:
I'd provide direction for a section ("Let's start with a thought experiment asking the reader to recall a favorite meal")
Cline would generate a draft based on my outline, style guide, and previous examples
I'd edit the draft, reshaping sections that didn't match my voice or adding personal observations
We'd refine each section until it flowed naturally into the next
Another key aspect of refinement was injecting my personal voice and specific phrasing. Here's an example of how I guided Cline to incorporate a joke about "training datasets" in the first thought experiment:
Me: One edit I'd like to make. Talk about how you scan through memories of memorable meals (your training dataset). i.e. I'd like to make a joke about how these memories are our training dataset. Another comment I'd like to make is everyone's training sets are different. If you're like me and don't eat out then, the scan takes less time.
Cline: (After implementing the changes) "What happened in your mind just now? You likely scanned through memories of memorable meals—your personal 'training dataset,' if you will—weighing them against each other until you settled on one that stood out... And just like AI models, everyone's training datasets differ. If you're like me and rarely eat out, your search might be quicker but more limited in scope."
This shows how I can provide very specific instructions—down to the exact phrasing of a joke or a personal aside—and the AI can seamlessly weave it into the existing text.
Final Touches and Publishing
Crafting the Perfect Title
After about 1.5 hours of work, the article was ready. I added a title and subtitle, reviewed the piece for consistency, and loaded it into Substack. The title selection process itself was another collaborative effort:
Cline: (After suggesting several title options like "The Search That Feels Like Creation" and "Navigating the Library of Babel")
Me: I wonder if we should be explicit like "Can AI be creative?" I want to hit the nail on the tension people are feeling that AI is mathematical. Like "Can a language model be creative?" and the subhead could be "The mechanics of creation and how AI creates a lot like us"
Cline: I like your direction of addressing the AI creativity question more directly. Here are some refined options:
Can AI Really Be Creative? / Subhead: Why the mathematics of language models mirrors human creativity...
Can a Language Model Create, or Just Search? / Subhead: The surprising mechanics behind both AI and human creativity...
(Provides several more options)
This exchange highlights how AI can rapidly generate alternatives, helping to explore different angles. My role was to provide strategic direction—identifying the core message and emotional chord I wanted the title to strike. The AI then adapted its suggestions to that new direction. The final title, "Search or Create: The Hidden Mechanics of Thought," emerged from this iterative process, where AI suggestions helped clarify the core theme I wanted to convey.
In my next article, I'll dive deeper into the lessons I've learned over two years of weekly publishing, how my approach has evolved as models improved, and where I believe this system is headed.
This journey of working with AI has been full of discoveries and continuous adaptation. The system I use today barely resembles what I started with, and I expect it will transform in the weeks and months ahead.