AI-generated YouTube thumbnails for $0.12 using a five-step prompt system

Executive overview

Creating thumbnails with AI is cheaper and often higher quality than hiring a designer. The bottleneck isn't the image model — it's writing prompts that produce usable results.

This five-step system solves that by having AI research best practices and write its own image prompts, leaving you to provide only the video title and notes.

The core insight: AI-researched prompts outperform human-written prompts because the model knows what the image model needs — you don't.

The five-step process

1. Draft a base prompt covering what (task), why (intent), and how (constraints)
2. Run the base prompt through a prompt improver (Claude or OpenAI console) to inject model-specific best practices
3. Feed the improved prompt into a thinking model (e.g. Claude Opus) to research best practices and generate optimised image prompts
4. Run those prompts through Imagen 3 (Gemini app or Google AI Studio) to generate thumbnail variants
5. Remove watermarks if using the Gemini app (e.g. Canva magic eraser)

Writing the base prompt

• Answer three questions: what (the task), why (the intent), how (constraints)
• Giving the AI the why lets it infer unstated requirements
• Key constraints to embed: high contrast foreground/background, one or two focal points, legible at phone scroll speed, authentic rather than overdramatic
• Specify two research areas: (1) best practices for the target channel or platform, (2) best practices for prompting the specific image model — both as of today
• Provide two runtime inputs: video title and transcript or notes
• Instruct the AI to return multiple prompt variants suitable for A/B testing

Using the prompt improver

• Claude: console.anthropic.com → Generate prompt
• OpenAI: platform.openai.com → Prompt optimizer (select target model, e.g. GPT-4.1)
• Output is longer and more structured — the model knows where to go and what to produce
• Do this once and embed the improved prompt in a Claude or GPT project; reuse it by dropping in title and transcript each time

Choosing where to generate images

Gemini app

• $20/month for consistent access (~100 images/day); free for occasional use
• Watermark on downloaded images — removable with Canva magic eraser or similar tools
• Simpler interface

Google AI Studio

• Pay-per-use: ~$0.10–$0.12 per image at 2K resolution; ~$0.20–$0.25 at 4K
• No watermark
• Requires API key linked to a billing account in Google Cloud Console

Tips for better face-matching in thumbnails

• Provide four or more reference photos of yourself at different angles and expressions
• Multiple reference images improve likeness accuracy over a single photo
• Embed in the base prompt that your face must appear in every variant