The generic AI thumbnail generator is easy to beat. Instead of asking one model for a picture, you split the job: Claude writes the concept and the image prompt, Gemini draws it, and a short rules document keeps the composition consistent with YouTube packaging principles.
What This AI Thumbnail Generator Actually Does
This AI thumbnail generator takes a working title, a short description, and an optional reference photo, then returns three distinct 16:9 thumbnail mock-ups styled like YouTube feed tiles. Each card has a download button and a two-to-three sentence explanation of the psychological hook behind the concept.
The tool was built by Ahrefs' presenter inside Agent A, Ahrefs' AI agent workspace that can query Ahrefs SEO data and build small applications. The presenter describes it as trained on eight years of YouTube packaging practices gathered from paid consultants, encoded into a document uploaded alongside the build prompt.
The presenter is explicit about the intended use: the output is for prototyping. The generated mock-ups go to a human thumbnail designer, who uses them as direction rather than final artwork. Faces are handled by uploading a reference photo, and the generator adjusts expression and composition around that person.
The Build Prompt: Two Models in One Pipeline
The pipeline splits creative planning from image rendering. Claude, Anthropic AI assistant, generates three thumbnail concepts and writes a detailed image prompt for each. That image prompt is passed to Gemini, Google's image-generation model, which produces the actual 16:9 image with the text overlay baked in.
The build prompt, as the presenter reads it on screen, specified:
- An internal console application for the team
- Three distinct thumbnail concepts per run, rendered as downloadable YouTube-feed-style mock-ups
- A two-to-three sentence rationale per concept covering the psychological hook
- Inputs: a working title, a script or short description (the presenter found a short description works better), and an optional reference photo of a person
- Output: three thumbnails side by side, expandable why-it-works cards, and a permanent history tab
A key instruction was to upload a best-practices document and require Claude to generate concepts that follow it. The presenter notes the tool behaves much like Claude own agent mode because the build ran on a Claude model; the exact model version changes as Anthropic updates its lineup, so check the Anthropic model page for the current generation.
Which Model Does What in the Pipeline
The three components have distinct roles, and keeping them separate is what makes the output consistent.
| Component | Role in the pipeline | Owner |
|---|---|---|
| Claude | Writes concepts, image prompts, and rationale | Anthropic |
| Gemini | Renders 16:9 images with baked-in text | |
| Agent A | Hosts the app, plans the build, queries SEO data | Ahrefs |
This separation matters for reproducibility. If thumbnails come out generic, you adjust the concept-planning side (the Claude prompt and the best-practices document). If rendering quality is the problem, you adjust the image prompt format sent to Gemini.
The YouTube Best Practices Baked Into the Prompt
The training document distills years of packaging notes into rules the concepts must follow. The presenter walks through the visible ones in the generated examples:
- ### Three focal points maximum
The document limits each thumbnail to three focal elements. In the golf-course example, that is the presenter in white (contrasting with the background), the club, and the overgrown course. Overloaded thumbnails lose clicks because viewers scan, not read.
- ### Contrasting colors
The generator is instructed to use contrasting palettes, such as blue against green with a hint of yellow. Bright, high-contrast combinations survive compression at small feed sizes.
- ### Short text overlays
Long text blocks are explicitly avoided because viewers do not read them in-feed. The overlay text is baked directly into the generated image rather than added later.
- ### Close-up subject pointing at the subject matter
One preferred concept is a large close-up of the presenter gesturing toward the video's object, such as a sinking golf cart, giving the image a built-in curiosity gap.
Why Build It in Agent A Instead of Claude or ChatGPT
The differentiator is data access. Agent A has unrestricted access to Ahrefs data, so the same workspace that builds the generator can also research which video topics are worth packaging in the first place. The presenter demonstrated this as a combined video-SEO and YouTube-SEO workflow:
- Ask Agent A for three high-traffic digital-marketing keywords where YouTube videos already rank on Google's first page.
- It queried Keywords Explorer and pulled SERP data, working through 5,197 rows covering 200 keywords before returning three candidates.
- For each keyword it returned search volume, keyword difficulty, intent, and the estimated monthly visits to the currently ranking video.
- It proposed working titles, then ran them through the thumbnail generator automatically.
In the demo, the keyword "Facebook Ads Library" came back with informational intent and a ranking video in position eight reported at roughly 328 visits per month. The presenter noted the traffic figure looked like an overestimation, which is a fair caveat: Ahrefs' video-level estimates are modeled, not measured. He also read the traffic graph as a signal the video earns passive search views rather than browse-driven spikes, because views kept growing a year after publication.
The titles Agent A proposed mixed search and browse phrasing, for example "How to use Facebook Ads Library to find winning ads" alongside the exact-match "Facebook Ads Library." The presenter noted you could adjust the prompt to demand search-focused titles if that were the only goal.
Build Time and Cost, As Reported
The speaker reports a build time of roughly 16 minutes and a cost of $3.29 for the application itself. Both figures are his own first-hand account of one run, not an independent benchmark, and costs will vary with model pricing, prompt length, and how many clarifying questions the agent asks during planning.
During the build, the agent asked planning-style questions: whether to include a brand logo placeholder (declined), whether history items could be deleted (declined, to keep all versions comparable side by side), and a technical question about the Gemini integration. Answering these well is part of why the finished app matched the brief.
Limitations Worth Knowing Before You Build One
Three limitations come directly from the demo. First, likeness is approximate: the presenter said generated people "typically look like me," which is why he frames the tool as prototyping support for a human designer, not a replacement for one.
Second, concept quality varies. In the Tom Brady test, he rejected one of three concepts outright. Generating three at once is the design's answer to hit rate, not a guarantee that every output is usable.
Third, the keyword workflow depends on Ahrefs data, including modeled traffic estimates the presenter himself called an overestimation in one case. Treat the estimated monthly visits to ranking videos as directional. The underlying search-demand data is more reliable than any single video-level figure.
A related walkthrough of stack-based builds and AI tooling is available from Dev Doido do canal do youtube at crazystack.com.br.
Frequently Asked Questions
- How much did the AI thumbnail generator cost to build? The presenter reports $3.29 and about 16 minutes for the build inside Agent A. This covers one run; your cost depends on current model pricing and prompt length.
- Which AI models does the thumbnail generator use? A Claude model from Anthropic generates the concepts and image prompts, and Google's Gemini renders the 16:9 images with text baked in.
- Can the generator put a specific person in the thumbnail? Yes. You upload a reference photo, and the generator adjusts expression and composition around that person. Likeness is approximate, which is why the output is intended as a prototype for a designer.
- Why build it in Agent A rather than Claude directly? Agent A can query live Ahrefs keyword and SERP data, so it can find rankable video topics and feed them straight into the thumbnail generator in one workflow.
- Can I install the same generator myself? The presenter says readers with an Agent A workspace can install it from the apps panel, or rebuild it from scratch using the prompts shown in the video.
Turn Your Own Video Knowledge Into an Article
The lesson from this build is that packaging knowledge locked inside a workflow becomes far more useful once it is written down and structured, exactly what the best-practices document did for the thumbnail generator. If you have that kind of knowledge sitting in your own YouTube videos, you can give it the same treatment.
Skala Blog turns a YouTube video into a written article: paste the URL, the video is transcribed, and you get a structured draft you can edit and publish.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits