Billionhands · Social ranking product
Rankings people trust, drafted with AI
Billionhands turns opinions into rankings people create, publish and debate. I designed the product model behind it: where AI reduces effort, where the creator stays in control, and where the community gets a say.
Discovery by category, and interest selection during onboarding.
Results
From generating rankings to assisting the people who make them
Research changed the direction of the product. People trusted rankings with a person and social proof behind them more than lists generated by AI. The shipped model kept generation only where no person had ranked yet, as a draft the person who asked could claim and change.
- AI drafts the gapsSearches with no human ranking get a generated draft instead of an empty result.
- People frame the questionA ranking starts with a person’s title, candidates and order, for drafts and handmade lists alike.
- The crowd finishes itAfter publication anyone can add candidates and vote. Nothing is removed, and the order follows the votes.
The problem
A ranking is information, but it is also a point of view
The obvious AI feature was to let someone type a topic and generate a complete ranking. That solved speed, but it created a more important product problem: if the model chooses the candidates and decides the order, whose opinion is being published?
I reframed the challenge around authorship. We kept generation, but changed what it produces: a first draft the person has to claim, not a finished opinion. The system does the work that is expensive or repetitive. It does not quietly take over the choices that make a ranking personal.
Research
People trusted people, not generated lists
We interviewed more than 15 people while shaping the product. Rather than treating research as validation for an AI feature, I used it to understand the jobs inside ranking creation: finding material, expressing preference, and making something worth sharing with other people.
A person may know what they want to rank without having every candidate in mind.
Let the system help discover candidates and reduce blank-page work.
Framing the question, picking the first candidates and ordering them are not setup—they are the opinion. After that, the ranking belongs to the debate.
Keep those actions explicit and fully controlled by the creator.
A ranking only becomes a debate if other people can change its outcome, not just comment on it.
Let people add candidates and vote on every item. Keep the question, and the name of who asked it, fixed.
The interaction model
Separate discovery, judgement and social validation
The research led to a simple division of responsibility. The system is strongest at drafting and proposing material. The creator owns the ranking. The community reacts after publication.
Turn a prompt into a first draft, interpret broad intent and propose candidates.
Set the question, then add, reject, remove and order candidates until the first version represents their point of view.
Add candidates and vote item by item. The published order follows the votes. Nothing is removed and the question does not change.
AI draft
A prompt starts the ranking, the person finishes it
We still used generative AI, where it helped most: searches that no person had ranked yet. The home screen leads with one action, write what you want ranked. Billionhands looks for the best existing match and, when no person has made that ranking, drafts one and shows similar rankings beside it. That removed the blank page, the most expensive moment in making a ranking.
The draft is never presented as finished. It says how it was made, asks the person who searched to sign up and claim it, and invites them to change it and suggest what is missing before anyone else votes on it.
From prompt to draft: search first, then a ranking labelled as generated, with a way to save it and add what is missing.
- Labelled as a draft“Generated by you using the power of Billionhands AI” sits inside the list, where the person reads it.
- Saving means claimingKeeping the draft asks for an account, so every saved ranking has an author.
- Gaps invite edits“Missing something in the list?” and a sticky Add turn the draft into a starting point.
Creation
The model leaves before the opinion is final
Candidates enter through one search. Add, Adding, Added and duplicate states are the predictable base. Suggestions from the model go through those same states, so they can always be undone.
Once the candidates are in, the interface becomes intentionally manual. The creator removes, adds and drags items into order, then checks a preview before anything is public. This was the boundary that mattered most: AI could help assemble the material, but the first version, its candidates and its order, remained the creator’s work.
A ranking built by hand: title, candidates from one search, manual ordering and a preview before publishing.
The three designers worked to three principles. Model output never bypasses the selection states. Order is always set by people, first the creator and then the community, never the model. Generated and handmade rankings use the same components, so neither reads as second-class. That last one is why the ranking card, vote control and candidate states are identical in both paths.
Community
The creator asks the question, the community answers it
Once a ranking is published, it only grows. Anyone can add a candidate the creator never considered, and people agree or disagree item by item while the list re-sorts as they do. Nothing can be removed and the question cannot be changed. The creator can delete the whole ranking, but not undo what the community did to it. Their version is where the debate starts, not where it ends.
Every item also links outward. An item’s page shows the other rankings it appears in, so one opinion leads to the next debate.
A published ranking with votes per item, and an item page that leads to other rankings.
What I learned
The useful AI feature was a boundary, not a takeover
We started by asking how much of ranking creation AI could automate. Research turned the question around: people trusted rankings with a person behind them. AI earned its place where no person had ranked yet, as a draft someone could claim, and as search that saved work without making choices. Neither the list nor the order was ever the model’s: one person starts them, people finish them.
What I would do differently: put the claim step in front of people sooner. We validated that people trust human rankings, but not how much work they would do to adopt a generated one.

