How to Write an Effective Prompt
Concrete prompting strategies for the Quick and Advanced lanes — what to include, what to skip, and how the 4-block composition scaffold uses your words.
The two prompt modes
FatThumb has two generation engines, each with a different prompt contract:
Quick lane (structured engine): Your prompt fills the user-direction slot inside a 4-block composition scaffold: FOREGROUND → STYLE → SUPPORTING OBJECT → BACKGROUND. The template fills in the STYLE, SUPPORTING OBJECT, and BACKGROUND blocks with archetype-specific guidance. Your prompt supplies the subject and intent; the template handles the visual structure.
Advanced lane (freeform engine): Your prompt is the primary creative driver. The FOREGROUND identity-lock (if you have a Person attached) and the inspiration mood-board framing are added automatically, but you control composition, mood, and content through your own words. There is no forced supporting-object or background block.
What to include in your prompt
Be concrete about the subject. "A developer looking shocked at their screen" is better than "developer." The more specific you are about what the person is doing and expressing, the closer the output is to what you have in mind.
Describe text overlays if you want them. If your thumbnail should have a text callout (e.g. "I QUIT", "WAS WRONG", "THEY SAID NO"), include it in the prompt explicitly. Keep it short — 1–4 words — because long text in AI image generation is unreliable.
Name the mood or energy. Words like "dramatic," "high-energy," "calm," "tense," "celebratory" give the model tonal direction that affects color, lighting, and composition choices.
Reference the video's angle if it helps. "Reacting to a rejection email" tells the model something about the emotional context. "Tutorial on fixing a merge conflict" gives it a subject matter anchor for the SUPPORTING OBJECT slot in Quick.
What to leave out
Let the Person handle the face. If you have a Person profile attached, you do not need to describe the person's appearance in the prompt — "dark-haired man in a blue hoodie" adds noise and can conflict with the identity-lock. The Person photos are the authoritative face source.
Skip generic filler. "Professional, eye-catching, high-quality YouTube thumbnail" is the default assumption of every generation — naming it wastes prompt space without adding signal.
Avoid extremely literal scene descriptions. "Show a person sitting at a desk in front of a white wall with a monitor showing the VS Code editor open to a file called main.ts" is too constrained. AI image models work better with intent and mood than with pixel-level scene reconstruction.
Quick lane tips
In Quick, the prompt is short by design. Aim for 1–2 sentences that describe the subject and the video's core angle. The template does the heavy compositional work. If your prompt describes the subject well and the template matches your video style, the result is often strong on the first try.
Lean on the video project flow: if you paste a YouTube link at the start of a Quick session, the AI analyzes the transcript and pre-fills a concept. You can edit that concept before generating — it is much faster than writing from scratch.
Advanced lane tips
In Advanced, you can go longer. A 3–5 sentence prompt that covers subject, action, mood, color direction, and any text overlays gives the model rich signal to work with.
Use @mentions in the prompt to tag Persons, Inspirations, or Images inline. This is an alternative to the panel selectors and useful for precise ordering when you have several assets.
Combine a strong prompt with a Style Reference that has the right color grade or composition energy — the reference provides visual signal that words alone cannot reliably reproduce.