A Personal Story That Might Inspire Other Creatives
Honestly, if you’d told me just two years ago that I’d be building complicated apps, like my Mozaix system, which now has tens of thousands of lines of code, I would have looked at you like you were joking. For most of my life, coding felt like this weird, foreign language. Programming? That was for someone else, the kind of people who enjoy impossible math problems or dream of building rockets—not me.
So here goes my honest story about stumbling into GenAI for development, messing up a bunch, and slowly figuring out that maybe other designers and creatives could actually relate. If you’re nervous about coding or just dipping your toes into AI tools, I totally get it. Maybe hearing about my mishaps (some of them truly cringe-worthy) will help you skip a few of the pitfalls I ran into.
MacOS desktop showing a split-screen view of multiple terminal windows and a browser tab with ClaudeAI open. The terminal displays Python code related to a mosaic algorithm, while the Claude interface contains a detailed technical discussion on AI photographic mosaic matching algorithms, including mathematical cost functions and implementation strategies.

Collaborative code development in action: Claude Desktop and Claude Code in refining the mathematical framework of the Tsevis Neural Mosaic Algorithm, blending human intuition with AI-powered precision. This dynamic exchange highlights how large language models can act as co-engineers in complex creative coding workflows.

Here's the truth: my coding skills were basic. I copied scripts from the internet, breaking things as often as I fixed them. Automation tools and simple workflows felt like speaking a language I barely understood, like ordering pasta in broken Dutch and ending up with soup (true story). 
Before GenAI, programming felt like an exclusive club full of jargon. I was the person Googling "how to center a div" and still getting confused.
Then something magical happened. I started chatting with Claude about my creative challenges, and suddenly, I wasn't just copying and pasting code; I was actually understanding it. Claude didn't judge my beginner questions or make me feel stupid for not knowing basic concepts. It was like having a patient tutor who could explain complex ideas in plain English.
Everything changed when problems that once felt like towering walls suddenly became doors I could walk through. With Claude and other AI systems by my side, I could finally dive into scientific papers on color theory and unravel mathematical models that used to leave me stumped. These tools enable me to explore physics and interdisciplinary research far beyond what art school ever taught, sparking the creation of applications with capabilities I once thought were out of reach.
The Tsevis Deep Mosaic algorithm—sounds fancy, right? It's essentially a smart system that analyzes thousands of images and determines how to arrange them into coherent art pieces. It utilizes a model called ResNet50 (a type of quite old (2015) Microsoft Research AI "brain" for analyzing images) in conjunction with traditional color matching. Think of it as having a super-intelligent assistant who never gets tired of sorting through your photo collection and finding the perfect match for every tiny piece of your mosaic.
The DiSalvatore grid system—this one has special meaning for me. It's named after the late Nino Di Salvatore, founder of the Scuola Politecnica di Design in Milan, where I completed my master's in visual design. Di Salvatore taught "Science of Vision"—a Gestalt theory-based approach that completely transformed how I see, understand, and create visual work. The system uses facial detection and computer vision models (like OpenCV's Haar cascades combined with more modern approaches) to create mosaic layouts that respect the underlying image structure. My entire Mozaix application is filled with references to people I admire—professors I've learned from, as well as personalities in computer vision, AI, and design that I deeply respect. Instead of randomly placing images, this system understands where the important visual elements are and arranges everything accordingly.
These aren’t just simple scripts for resizing images. They’re powerful systems, robust enough to tackle the intricate projects I take on for major brands. And the best part? If someone like me—who once lost hours hunting for a missing comma—can build these, then you absolutely can too.
Two Years of "Wait, Did I Actually Build That?"
Looking back at my journey with Claude feels like watching someone else’s highlight reel. In the early days, it was like talking to a really smart friend who never got tired of my basic questions. However, those chats had limits—imagine trying to explain a big project but only being able to show three pages of notes at a time.
Then came the game-changers: expanded context windows (Claude could finally keep track of our entire conversation without forgetting what we were building), Project Knowledge (now simply called Files—think of it as Claude having access to my entire project folder), and most recently, Claude Code, which basically gave Claude the ability to actually implement the ideas we discussed.
With every update, it felt like my creative superpowers were leveling up. Claude started to grasp my half-formed ideas and turn them into real, working solutions. I could finally explain problems in my own language, not just in tech-speak. Suddenly, there were fewer maddening crashes and more jaw-dropping moments when everything just clicked.
When you’re creating work for clients like Google or the American College of Cardiology, you need tools that work at a professional level. Claude has kept up with my more ambitious projects, which is honestly impressive.
The Tsevis Mozaix application displaying a high-resolution mosaic in progress. The interface shows adjustable parameters like tile size, spacing, and matching methods, with a preview of a colorful David Bowie portrait being constructed from smaller image tiles. The settings panel includes advanced controls for color, edge structure, and semantic content weighting.

Fine tuning the TsevisDeepMosaic algorithm in Mozaix application, my digital mosaic creator. 

A Recent "Aha!" Moment (Or: How I Learned My Tools Have Personalities)
Last week, I had one of those moments that every developer knows and dreads: something that worked perfectly yesterday was suddenly producing garbage results. My TsevisSmart algorithm, which had been creating beautiful photomosaics, decided to throw a tantrum and start making art that looked like it was assembled by a colorblind robot having a bad day.
Here’s where it got both interesting and a bit embarrassing. The problem was a ComfyUI upgrade that broke my system dependencies, made worse by my habit of not using virtual environments. For anyone not technical, it’s like mixing all your art supplies in one box instead of keeping them organized. Sooner or later, the watercolors mix with the oil paints, and everything turns into a mess.
Pro tip from someone who learned the hard way: Always use virtual environments (`python -m venv project_env`) for your projects. My long-time friend See-Ming Lee—a talented designer and developer—had advised me about this years ago, but I was too lazy to set them up properly. Think of virtual environments as separate workspaces for each project—they prevent the kind of cascading disasters that had me questioning my life choices at 3 AM.
But here's the real revelation: I initially asked Claude Code to help, assuming it was my go-to tool for everything. Claude Code did what it does best. It dove deep into the algorithm logic, analyzed the code structure, and suggested improvements. It was like asking a master carpenter to fix what you think is a structural problem. Spoiler Alert: It didn’t work!
Then I tried regular Claude with the same issue. Completely different approach. Within seconds, Claude was like, "Hold up! If the same files worked yesterday, this smells like an environment issue," and started digging through my terminal logs. Found the smoking gun immediately: `TensorFlow attribute error: module 'ml_dtypes' has no attribute 'float8_e3m4'`.
To be fair to Claude Code, it suggested that there might have been an environmental change, but as it wasn’t specific, I couldn’t remember that I had clicked the Update ComfyUI button yesterday. 
It was like the difference between a detail-oriented craftsperson and a detective. Both brilliant, but with completely different superpowers.
Understanding Your AI Partners: It's Like Having Different Types of Smart Friends
This incident taught me something valuable about working with AI tools—they're like having different types of brilliant friends, each with their own strengths and blind spots.
Claude is your detective friend—the one who steps back, looks at the big picture, and asks the right questions. When you're stuck, Claude excels at:
- Playing Sherlock Holmes with your problems ("If it worked yesterday, what changed?")
- Connecting dots across multiple files and systems
- Explaining complex concepts in human language
- Planning out entire projects before you start coding
- Being your research buddy for topics you don't understand
Claude Code is your hands-on friend—the one who rolls up their sleeves and gets stuff done. When you know what needs to happen, Claude Code dominates at:
- Actually creating and editing your files
- Running commands and installing things
- Testing ideas quickly and giving immediate feedback
- Managing your project setup and configurations
- Being your implementation partner who never gets tired
Here's my workflow that's saved me countless hours of frustration:
Start with Detective Claude
When something's wrong or you're planning something new, begin here. It's like having coffee with a brilliant consultant who helps you figure out what you're actually trying to solve.
Perfect for moments like:
- "This used to work, why doesn't it now?"
- "I want to build something, but I'm not sure how to approach it"
- "My code is slow and I don't know why"
- "I'm getting weird errors that don't make sense."
Move to Builder Claude Code
Once you have a plan, Claude Code becomes your implementation partner. It's like working with a skilled craftsperson who can turn your ideas into reality.
Switch when:
- You know exactly what needs to be built or fixed
- You need to create or modify multiple files
- You want to test things and see immediate results
- You need to install or configure system stuff
Back to Detective Claude for Polish
Return to regular Claude when you need to step back and think strategically again.
Perfect for:
- "This works, but how can I make it better?"
- "What should I build next?"
- "How do I explain this system to someone else?"
- "Something feels off, but I can't pinpoint what"
Everything falls into place when you stop expecting one tool to be a magic wand and instead let each shine in its own way. It’s like knowing when to call your meticulous friend for the details and when to brainstorm with your visionary friend. Both bring something special to the table.
A vibrant digital mosaic portrait of David Bowie composed entirely of pixel art tiles. Each tile features iconic imagery from pop culture, video games, and retro aesthetics, arranged to form Bowie’s face in a stylized, colorful composition. The background is a deep teal, and the artist’s signature “Tsevis” appears at the bottom left.

Crafting a pixel art mosaic of David Bowie. Tesserae are created with various diffusion models on ComfyUI and assembled with my Mozaix software.

My Honest Experience with Other AI Coding Tools
Let me be honest about the other AI tools I’ve tried, since no one wants to read advice that sounds like a sales pitch. Please take my opinion as just a beginner's experience and with a grain of salt. Just know that I am not sponsored by Anthropic or any other company. I admire them all for their products and their technology. 
Gemini CLI and AI Studio are my nr.2 choice after Claude. The seem to me like that really smart friend who's great at quick help but gets frustrated when things get complicated. They're much cheaper (or free!), which is awesome for simple tasks, but I've noticed they tend to give up when problems get hairy. It's like asking for directions and having them say, "Just go that way," and walking off when you ask for clarification.
Qwen LLM and Qwen Code performs similarly to Gemini—solid for straightforward stuff, but when you hit a wall, it often just... stops helping. Qwen LLM is a great multimodal model that I love using for many daily tasks, coding included. 
ChatGPT and GPT-3 OSS (Open Source) models excel in basic implementations. I am constantly amazed by ChatGPT's writing abilities, and it never fails to impress me when I ask it to review my code or spark new ideas. The depth of Computer Vision knowledge in these models is both astonishing and, at times, delightfully unexpected. I have even orchestrated a unique conversation between ChatGPT and Claude, sharing code and ideas and inviting both to “collaborate” through my copy-pasting. The theoretical results are nothing short of groundbreaking, with concepts that sound truly remarkable. Yet, despite the promise, I have not yet fully brought any of these ideas to life. The complexity remains a formidable barrier. But I have learned so much. It’s like listening to some great academics debating your favorite subject. 
However, I would definitely encourage you to experiment in this way. 
DeepSeek is genuinely impressive for simple issues, especially considering the pricing. For me it's a precious assistant when I want to save some precious tokens from Claude. Whenever I have to edit a complicated method or ask to merge code or just check the correctness of some expression DeepSeek is there to help me.
Inception Labs Mercury and other revolutionary specialized tools each have their moments of brilliance, but here's what I've learned through many late-night coding sessions: when you're working on something complex and professional-grade (like creating visual systems for any serious professional project), reliability isn't just nice-to-have—it's everything.
Other tools often make syntax errors that take several messages to fix, or lose track of the project halfway through. Claude, on the other hand, stays accurate and focused during long problem-solving sessions. It’s like working with a colleague who pays attention the whole time, instead of one who keeps getting distracted.
If This Sounds Impossible, I Felt the Same Way
If you’re thinking, “This sounds amazing, but I could never do that,” trust me, I thought the same two years ago.
Start with your daily annoyances. That task that drags forever, that messy workflow, that “there has to be a better way” moment. Those frustrations are the doorway to building solutions.
Here’s the secret: you already know the hardest part. As a creative, you see what good work looks like. You know what details matter. You are the user, which makes you the best guide for AI.
Don’t just chase quick fixes. Ask bigger questions: What do trusted sources and real research say about this problem, and how can I build from there? Easy answers have their place, but the real value is in conversations that spark ideas you wouldn’t reach alone.
I still remember when one of my apps processed a thousand images in minutes. Or the first time my Mac “saw” dozens of parameters in a mosaic tile, work that would’ve taken me days. I just stared at the screen thinking, “Did I really build this?” That feeling never gets old.
The Weirdest Part: It's Like Your Brain Gets Upgrades Too
Here's something nobody prepared me for: working with increasingly sophisticated AI models actually changes how you think about problems. It's as if your brain is receiving software updates alongside the AI.
When I collaborate with newer, more powerful AI models, I find myself approaching challenges differently. Problems that once seemed impossibly complex now feel approachable. Technical concepts that were foreign become familiar. It's not just that I have better tools—I've become a better problem-solver.
This sparks a powerful feedback loop: better AI partners help you tackle larger challenges, which sharpen your skills and unlock even more ambitious projects. With every step, your creative abilities expand.
A multitasking desktop environment featuring multiple windows of Internet browsers open in ChatGPT and Claude.

Developing the ultimate mosaic algorithm becomes a dialogue: ChatGPT talks to Claude and vice versa.

The Real Revolution: Quality, Not Just Speed
Saving time with automation is beneficial, but that’s not the most significant change. The real benefit lies in raising the quality of your work and creating things that were previously impossible.
My applications don't just speed up existing processes; they enable entirely new kinds of creative expression. Advanced color harmony analysis, perceptual matching algorithms, and neural network-guided composition—these capabilities open creative territories that didn't exist with traditional tools alone.
When I'm working on complex medical visualizations for pharmaceutical companies or dynamic visual content for scientific publications, the quality requirements demand solutions that push beyond conventional software limitations. AI collaboration makes that level of creative sophistication accessible to individual artists and small studios—not just big tech companies with massive development teams.
Welcome to the Creative-Developer Renaissance
We're living through something special: the emergence of creative-developers. Artists who code their visions into reality. Designers who build the tools they need. Visual communicators who engineer their own creative possibilities.
This isn’t about replacing your creative skills—it's about amplifying them exponentially. Every creative professional who utilizes AI-assisted development is building an advantage that continues to grow.
Two years ago, I was intimidated by code. Today, I'm building applications that enable me to create work I never thought possible. The distance between those two points? Thousands of conversations with an AI partner who never made me feel stupid for asking basic questions.
If someone like me—who once spent an embarrassingly long time trying to figure out why my code wasn't working (it wasn't saved)—can build professional-grade applications, then honestly, what's stopping you?
The future belongs to creators who combine artistic vision with technical skills. With AI as your patient and helpful coding partner, that future isn’t far off—it’s already here.
The only question is whether you'll be part of building it.
Charis Tsevis, August 2025
PS: I am just a visual designer and artist. Better read from the source: Claude Code: Best practices for agentic coding.
Back to Top