What AI Cannot Draw

The service needed a logo.

I'd had AI write the code, so why not have AI make the logo too? I fed Gemini a prompt: "Create a logo for a SaaS product called StackTube. Combine a YouTube play button with knowledge layers."

A result appeared. It didn't look bad. Clean, decent colors, the kind of shape you'd nod at if someone said "SaaS logo."

But something nagged. I couldn't articulate exactly what.

I changed the prompt and generated three more. Different directions each time. Abstract monogram, pipeline metaphor, lettermark substitution. Six total. I sent them to the control tower: "Give me your opinion as a senior logo designer."

The control tower's answer was not what I expected.

"The core problem isn't the direction. It's the limitation of the tool."

AI image generators produce "images that look like logos," it said. The things an actual logo requires—letter spacing, weight differentiation, pixel-level alignment, vector output—none of these can be controlled. That's why all six felt "something's off but I can't tell what"—a persistent awkwardness.

This was unexpected.

When making code, AI had been remarkably precise. Creating files, executing them, fixing errors. But when making logos, that precision didn't function. Code has right and wrong. Logos don't. Instead there's good and awkward, and that judgment comes from fine-grained control at the pixel level.

The control tower offered an alternative. Build it directly in SVG code, it said. Unlike AI image generation, SVG is code—every letter, every shape can be precisely controlled. You don't need to know code yourself; the control tower writes the SVG, you look at the result and adjust.

From here, the familiar pattern returned. Control tower plus builder workflow. "Widen this letter spacing a bit." "The triangle should feel like it's replacing the T's vertical stroke—right now it looks placed on top." "Attach the BETA badge to the upper right."

Then came another pivot, this time on color.

The AI's initial suggestion was teal. A color commonly used in SaaS logos. Clean, technical impression. But when I sent an actual screenshot of the site, the control tower said: "The site has a warm cream-beige tone. Teal doesn't match. You should use the amber from your CTA buttons."

Before the screenshot, AI had judged by generality: "SaaS typically uses colors like this." After seeing the actual site: "This site needs this color." Same pattern as EP.08. AI judges only within the context you provide. Add a screenshot as context and the diagnosis changes. With the payment error I'd shown claude.md to correct premises; with the logo I showed a screenshot. Different forms, same principle.

The final logo became a wordmark. "Stack" in bold, "ube" in medium weight. The capital T's vertical stroke replaced by a play triangle—a small design with deliberate intent. Amber. It sat naturally against the site's warm tone.

I discarded all six originals. But making those six built up a sense of "not this"—and that sense was necessary to identify the right direction. Without the discarded ones, I wouldn't have recognized the right one.

Asking AI to code and asking AI to design. The same prompt-result-revision cycle, but it operates differently. With code, AI can verify on its own whether something is right or wrong. Run the tests. But with design, AI cannot judge on its own whether something is good or awkward. That judgment comes only from a human eye.

So for someone building a service through vibe coding, design is actually the domain requiring more human involvement than code. Code has wide room for "handle it yourself." Design has wide room for "I need to look, feel, and say 'not this.'"

You can see even if you can't make. And in this case, seeing mattered more than making.

---

🔧 Technical Terms in This Episode

SVG (Scalable Vector Graphics) A format that represents images as code. Regular images (JPG, PNG) are collections of pixel dots that blur when enlarged; SVGs are composed of mathematical shapes that remain sharp at any size. The preferred format for logos.

Wordmark A logo form where the text itself is the logo. No separate symbol or icon—the brand name's font, color, and arrangement alone constitute the logo. Google, FedEx, and Coca-Cola are classic wordmarks.

Vector A method of representing images through mathematical relationships between points and lines. The opposite is raster (pixel-based). Vectors maintain quality at any size, making them ideal for logos, icons, and print materials.

Kerning Adjusting the spacing between individual letters. Even in the same font, certain letter combinations (e.g., AV, To) look visually uneven, requiring manual adjustment. AI image generators cannot perform this adjustment.

Favicon The small icon displayed in a browser tab. Typically 16×16 or 32×32 pixels. A logo must remain recognizable at this size, so a simplified favicon version is created alongside the main logo.