The 'Looks Too AI' Debate

The most persistent design conversation in the group: AI can produce a visually impressive flyer in a minute, but a trained eye spots it instantly. Why? And what does it take to make a flyer that looks intentionally human-designed?

The position split

Pro-AI-for-flyers (most members):

• "AI can provide the base; humans polish." Time-poor shluchim get a flyer in ten minutes, not ten days.
• "If hiring a designer costs $600 and AI costs $0, the economics are not subtle."

Skeptical (a vocal minority):

• "Telltale AI flyer. The design is busy and all over the place. It might be easy to produce, but it's not good graphics."
• "People want to hear you. They see AI, they see it's not you."
• AI graphics should not replace human designers any more than AI text should replace human writing.

The 7 AI Tells

The canonical taxonomy, distilled from the community master prompt and refined over many threads. See Designer Prompt for the full system prompt.

1. No intent. AI makes everything "nice," but nothing purposeful. Counter: one prompt equals one visual idea, statable in a single sentence.
2. Default symmetry. AI centers everything. Counter: asymmetry is the default; symmetry must be earned (a formal gala, solemnity).
3. Everything shouts equally. All elements carry the same weight. Counter: one dominant read, one secondary, some things deliberately quiet.
4. Fake type space. AI leaves generic poster space. Counter: design around the actual text, with exact placement and size ratios.
5. Global polish. AI smooths everything uniformly. Counter: specify imperfect directional lighting, material-specific textures, and at least one deliberate flaw.
6. Stylistic mush. AI blends styles. Counter: name one design tradition (Swiss typographic, grassroots broadside, editorial food magazine, luxury invitation, brutalist, duotone, Japanese minimalist).
7. Dead space vs. tense space. AI either fills every pixel or leaves lifeless voids. Counter: empty space must have tension, a relationship to the other elements.

The Hebrew and RTL problem

Through 2025 image-generation models garbled Hebrew letters and reversed sentence order, so the group always fixed the text by hand. As of mid-2026 the top models render short Hebrew reliably: GPT Image 2 preserves right-to-left flow and stroke order, and Nano Banana Pro passes an independent Hebrew-rendering eval. So verify the output, and keep the two patterns below as a fallback for wrong letters or for long or dense Hebrew, where accuracy still slips:

1. Strip and replace. Generate the flyer, tell the model to remove the Hebrew letters without touching anything else (leaving placeholders), then type the Hebrew in Canva yourself.
2. Background only. Generate the visual background in Nano Banana or Gemini, then add all text in Canva.

For typeset documents (LaTeX or Typst source, not image generation), Claude still needs reminding to write RTL and keep sentence order; it is better than GPT for that task but still imperfect.

The workflow that actually wins

A pattern that recurs across many threads:

1. Generate the prompt with one model. Ask ChatGPT to write a detailed image-gen prompt including your brand colors, audience, and event details.
2. Generate the image with another model. Paste it into Nano Banana (Google AI Studio) for the strongest visual output.
3. Edit text and details in Canva. Magic Layers decomposes flat AI images into editable elements.
4. Optional: bring the result into Adobe Firefly or Canva as editable layers.

A template prompt for redesigning an uploaded flyer:

"Redesign it with a strong visual upgrade while keeping the core content intact. Modernize the layout, improve typography, spacing, and alignment, and create a clear visual hierarchy. Refine the color palette for a more cohesive and striking look. Reduce clutter, enhance readability, and add subtle design elements where needed."

Open problems

• Generating multiple consistent images for a series. Workaround: describe a reference image in JSON detail (via Google Lens), then paste the JSON into each successive generation.
• AI cannot see what it generated, so edits drift back toward defaults.
• Chabad-specific iconography fails reliably: menorah arm count, Luchos shape, kippah and long-sleeve consistency. A "Chabad image skill" custom Gem is in progress.

Related

• Designer Prompt, the full system prompt.
• Flyer Design, the workflow page.
• Nano Banana, Canva, and ChatGPT in the Toolbox.