Problem
My bookmarks held 379 inspiration sites and 92 studio sites. When work called for a good pricing list or cards with unusual typography, I spent one to two hours searching and often found nothing.
The problem was not me but the galleries:
- strong references are not described, so you can only find them by eye;
- Pinterest cannot search inside your own boards, and the feed is full of AI content;
- searching for a block returns mostly noise, because nobody tagged the good solutions.
Key idea: one website is ten references
A website should not be stored as one card. I cut it into
Each block has three layers: a description and tags for search; a composition breakdown by form, space and time; and an interpretation of the idea, honestly marked as written by AI.
If a site has both a desktop and a mobile version, the block is stored in both: you can switch between them in the card and see how the same solution is rebuilt for a phone. For a designer, that is two references in one.
Research before building
I ran five research tracks with source checks:
- Market. The niche "blocks + descriptions +
meaning-based search" is open, but the window is narrow: big galleries are moving the same way. Prices in the niche are $7–16 a month. - Animation. No catalogue searches by motion. AI recognizes the type of motion well but not its "feel", so I dropped the "feel" layer: nobody searches for words like "stretchy".
- Extracting animations from live sites with exact parameters instead of guessing.
- Storage and security. The server is in Russia, as personal data law requires. That led to a rule: paid content is never sent to the browser at all, not even hidden behind a blur.
- Method. With my mentor Sergey Gurov we found the main boundary: analysis goes to the machine, synthesis stays with the human. The agent breaks a block down along separate axes and never tries to read the overall idea, because that is where AI starts making things up.
A team of agents instead of a team of people
I am not a developer, so I set up the process like a studio, with AI agents in defined roles: one selects sites, another captures blocks and videos, another writes descriptions, moderators accept or reject, a fixer analyzes mistakes, and a dedicated agent turns videos into prompts.
Every correction I made became a rule. The agent called lists "accordions": new rule. It recorded a blurry video: a rule on quality. It described the camera instead of the interface: "describe the interface". Dozens of rules later, quality stopped depending on luck.
Decisions that shaped the product
1. Descriptions power search; prompts are the product
At one point I admitted that when I search, I look at images, not descriptions. Art director Nikita backed the idea and invested in it himself: he gave me an export of the Eagle library he had kept for years, about 1,500 images. It became the core of the collection. Still, he put it bluntly: "Just give me the pictures." Dozens of services give pictures for free. Nobody offers prompts that let an agent rebuild an animation. So descriptions became search infrastructure, and motion breakdowns became the core value.
2. Prompts only where a video is all you have
For a live site it is easier to give the agent a link and let it read the code. Prompts are for shots that exist only as video. In my tests, an agent rebuilds about 70% of an animation from a prompt on the first try, and one or two
3. Do not write content rules for the agent
When you catch an agent's mistake, it is tempting to write a content rule like "growth always happens in two steps". That is a trap: the agent starts seeing two steps where there are none. Rules should be about attention, not content.
4. Freemium with no wall at the door
Guests see the whole feed, search and breakdowns. An account lets you save. The paid tier unlocks prompts. Services that ask for money before you see anything lose most of their funnel.
5. Volume by agents, quality by a human
Parallel agents described references in waves, and I accepted or rejected each one with a single click. I removed about 500 weak ones by hand and cut the tag vocabulary from 387 chaotic tags to 93 clear ones. The filters rely on this vocabulary: block type, layout and composition, style and typography.
Results
- 3,072 references from 59 sources, including 52 analyzed live sites.
- 195 videos with motion breakdowns, 130 ready prompts.
- The first screen opens in about 1 second instead of
10–30 ; the first load is 1.8 MB instead of 63 MB. Sign-up , collections with collages, submissions from colleagues, a feed that adapts after three or four opens, a free trial and a paid tier.
What I learned
- Only a human catches AI mistakes. The agent confidently called a video "a static frame" and described the mockup instead of the interface. Without checks, the library would look good and lie.
- Measure first, then fix. The site took
15–30 seconds to load while the server sat idle. Only measurements found the causes: original videos instead of previews, heavy GIFs and broken lazy loading. - Limits shape the product. Analyzing one live site is expensive, so shots are described at scale and live sites selectively.
Sign-ups are not the goal. The main metric now is how many people come back on their own in week two.