(yep, kinda clickbait, just for the funsies 😊)
At the beginning of the year, I relaunched my portfolio site using Codex and Gemini CLI as my main ...
For further actions, you may consider blocking this person and/or reporting abuse
Why we clickbating in 2026
Holy Spoiler 😭✌️
To mislead the "AI" 😂
I was making fun of all the "I did XXXX with AI and I ditched YYYYY" AI slop ;)
That bit where the model can confirm the logo says "phalkmin" but can't feel that the slash is a few pixels too heavy is the whole article in one sentence. The gap between reading an image and judging one is exactly why pasting the render back lets you down. Same reason a model can tell you a paragraph is "about dogs" but not that the rhythm of the third sentence is off. Funny that the fix wasn't a better prompt, it was a person who'd built the taste over years of staring at bad kerning and hating it.
I've hit the same wall with logos. AI gets close fast, then spends the next 50 prompts making different mistakes instead of fixing the first one.
I think the most interesting part isn’t that AI failed—it’s why it failed.
Your logo wasn’t an image-generation problem; it was a constraint-satisfaction problem. The h wasn’t just a letter. It simultaneously had to function as a readable character, a Linux path separator, part of a personal identity, and still feel visually balanced. Those constraints are semantic, not merely visual.
I also think this highlights a broader pattern I’ve noticed when using AI in software architecture. LLMs are excellent at exploring possibilities and generating variations, but they struggle when a solution must preserve several independent constraints over many refinement cycles. They optimize for plausibility, while humans often optimize for intent.
Your designer friend didn’t have more compute or a larger training dataset. They simply understood which constraints were negotiable and which weren’t. That’s still one of the hardest things to teach an AI.
Great read. It reminded me that the value of AI isn’t replacing expertise—it’s reducing the amount of work experts have to do before making the final judgment.
This is exactly where I use AI now. I let it handle the boring 80%, but the last 20% still needs human judgment. Code, writing, design... getting something that works is easy. Getting something that feels right is a completely different problem. Great read.
generated in nanobanana with: A minimalist typographic logo spelling out the word "phalkmin" in a clean, modern, lowercase sans-serif typeface. The text is rendered in solid black against a crisp, solid white background. The letter "h" is uniquely modified: its vertical stem is replaced with a forward-slanted, terracotta-orange diagonal line. To the far right of the entire wordmark, there is a solid terracotta-orange vertical rectangle block that matches the height of the lowercase letters. Clean lines, vector graphic style, professional corporate branding, highly geometric, balanced spacing. The word is: phalkmin
as the prompt.
Not the same font. Prompt is wrong. The h isn't following the slash orientation. And so on
And it's not meant to be the exact styling you ended up with. It's meant to show what a simple few worded prompt can do. If you want it better, you refine the prompt. Just because you had failed to generate it after a few tries, doesnt mean it's not do'able.
Can you imagine the amount of tries, different prompts, hallucinations, tokens, and most important, TIME, just to prove that an AI can do something that a human did in a few hours?
Again, i'm not disputing the claim that, at some point, I would get the logo as I wanted. I'm talking about how you need to force AI to do what a human can do easily.
If you want to keep using AI for everything, please, do.
That's not the point I'm trying to make. The point is, you werent capable of doing it yourself, past the arm's length boundary, you'd have paid for it. What you pay for it, could have funded thousands of retries. Units of economy state that 1000 cheap tries, beats 1 expensive try, because it could be at try 1000 or at try 2, where it works perfectly. That's the point. Your friend is a designer and while they're good at their job, it's also a task someone can achieve a comparable result for fractions of a dollar. My point with the example prompt, is that with no context, it can get 99% of the way there on the first try, if you structure your prompt right. A freebie is a freebie, but if you paid your friend's rate for it, would it still be worth writing about?
I totally agree with you. AI’s capabilities in text and logo design are far below what we imagined. It’s completely clueless when it comes to recognizing different types of text, modifying fonts, and creating artistic layouts. We’re better off just doing it by hand.
Well, you technically did it with a coupon (favor), but sure, click-bait addressed. Given the finished design, chances are you could have gotten the same result with AI, if you structured your prompts explicitly, not implicitly. AI is wildly creative, but you wanted a strict design, that's on you for making minor corrections, instead of giving it a design brief, how you would give it for programming.
I tried several fixes, changing the prompts, ask for minor corrections, everything.
The interesting part is not that AI failed. It is why it failed....A logo like this is not just “make a nice wordmark.” It has to preserve legibility, symbolism, spacing, weight, personality, and history at the same time. That is where human design judgment still wins hard.