Public-Safe Image Generation for Knowledge Bases

Generated images can make a knowledge base feel polished, but they are not automatically safe. A model can invent readable labels, fake logos, private-looking dashboards, or off-brand mascots.

Official Hermes Agent docs: https://hermes-agent.nousresearch.com/docs/reference/tools-reference

Prompting standards

• Ask for abstract shapes, icons, and panels instead of screenshots.
• State that there should be no readable text, letters, logos, secrets, or private UI.
• Keep exact article titles and labels in HTML/CSS, not inside the generated image.
• Use the approved brand palette and avoid invented mascot redesigns.
• Generate only as many images as the release can actually inspect.

QA checklist

• Open the actual image, not just the returned URL.
• Check for readable or fake text.
• Check for logos, watermarks, account names, or private-looking screens.
• Check crop, aspect ratio, contrast, and repeated hero-image use.
• Save with a descriptive filename under the public asset folder.
• Add alt text and a caption that explain the image without over-promising.

Pitfalls

• Trusting the prompt instead of the pixels.
• Publishing a hero image with garbled pseudo-text.
• Using an image provider the user did not approve.
• Letting generated visual metaphors imply a feature exists before it does.
• Reusing the same hero across unrelated pages.

Verification steps

• Inspect the image visually before referencing it.
• Confirm file size and dimensions locally.
• Load the article route and verify the image appears correctly on desktop and mobile.
• Replace or defer any image that fails public-safety QA.