YouTube Thumbnail

Create a clear 16:9 thumbnail built around one idea. Favor hierarchy, curiosity, and an instantly readable subject over visual noise.

Core Rules

• Use a 1280 x 720 canvas unless the user requests another size.
• Communicate one idea with one dominant focal point.
• Prefer no text when the image communicates the idea. Otherwise use one short phrase, normally one to four words. Do not repeat the full video title.
• Keep colors natural and controlled. Do not globally oversaturate the image, crush shadows, clip highlights, or push skin tones toward orange or red.
• Preserve photographed people. Do not regenerate, beautify, reshape, repaint, relight, recolor, sharpen, smooth, or otherwise alter their face, skin, hair, body, expression, clothing, accessories, or silhouette.
• Allow only cropping, uniform scaling, positioning, and a natural drop shadow around a person. Do not add strokes, keylines, glows, halos, rim lights, or bright outlines around people.
• Never invent or replace a recognizable person when a source photo is provided. Use the original pixels.

Workflow

1. Inspect the title, topic, source images, brand assets, and any reference thumbnails. Identify the single visual promise of the video.
2. Select the simplest composition that communicates that promise:
   • one subject plus a contextual background;
   • one subject plus one supporting object; or
   • one strong object or scene with optional short text.
3. Establish hierarchy at small size. Make the main subject recognizable before adding secondary elements. Keep important content away from the bottom-right timestamp area and allow breathing room near every edge.
4. Lock any source person before generating or editing other elements. Generate the background or supporting graphics separately, then composite the original person over them. If exact pixel preservation cannot be guaranteed, do not send the person through a generative edit.
5. Add text only if it supplies information the image cannot. Use one bold, legible type style with strong contrast and minimal effects. Avoid stacked captions, tiny labels, punctuation clutter, and multiple competing fonts.
6. Apply restrained finishing. Use selective contrast and color separation instead of extreme saturation. For a person, use only a soft, neutral drop shadow with modest opacity and blur.
7. Export as PNG, JPEG, or WebP at 1280 x 720. Preserve a lossless working file when the workflow supports layers.

Image Tool Guidance

• Inspect every supplied source image before editing it.
• Use image generation for new backgrounds, objects, textures, or complete thumbnails that contain no supplied person.
• Use non-generative compositing for supplied people when feasible. Keep the original person layer unchanged apart from crop, uniform scale, position, and drop shadow.
• When writing an image-generation prompt, explicitly request a clean 16:9 composition, a restrained natural palette, realistic contrast, negative space for optional text, and no embedded typography unless the tool is meant to render the final text.
• Do not claim a person was preserved unless the final image was compared with the source at full size.

Visual QA

Before delivery, inspect the full-resolution result and a small thumbnail-sized preview. Confirm all of the following:

• The concept is understandable within roughly one second.
• The thumbnail contains one clear focal point and no unnecessary decoration.
• Text is absent or limited to one short, readable phrase.
• No person has changed relative to the source.
• People have no outline, glow, halo, rim light, or edge color; any separation comes only from a natural drop shadow.
• Saturation and contrast remain controlled, especially in skin tones and brand colors.
• The subject and text remain legible at small size and are not obstructed by the timestamp area.
• The final file is exactly 16:9 and has no accidental crop, padding, artifacts, or embedded watermark.

If any check fails, revise the thumbnail and inspect it again before presenting the result.

Delivery QA

Do not treat an internal sandbox path, generated preview reference, or claimed export as a delivered file. Before finishing:

1. Confirm that the final image exists at a user-accessible local artifact path for the current task, using its configured output directory when available.
2. Verify the actual file type, dimensions, and size. When the file was copied from another local source, compare hashes to prove the artifact is complete.
3. Present the verified artifact with an absolute local path that the current Codex client can render or download.

Treat the task as partially complete and explain the blocker if the image is visually finished but cannot be exposed as an accessible artifact.