I wanted to build something that actually helps people fix their dating profiles. Not tips. A proper process.
I used Claude to run parallel research agents across the academic literature on dating app behavior — signaling theory, self-presentation, deception linguistics, mate selection economics. The agents pulled in findings from about 45 papers, and I sorted them into three tiers: peer-reviewed, platform data (Hinge/OkCupid's own numbers), and conventional wisdom. The research doc in the repo is honest about which is which.
Some findings that shaped the design:
Vague language is a linguistic marker of deception. People who lie in profiles use abstract, generic phrases. People who tell the truth use concrete, specific language. (Toma & Hancock, 2012, Journal of Communication)
Desirability on dating apps follows a power law. The top profiles get 10-100x the attention of the median. You cannot out-volume this curve. (Bruch & Newman, 2018, Science Advances)
"I have a great sense of humor" is what signaling theory calls a cheap signal — anyone can say it. Actually being funny in your prompt is a costly signal — you need the ability. Costly signals are processed as more credible. (Donath, 2007, building on Zahavi's handicap principle)
Creative, original profile text independently predicted attractiveness ratings beyond photos. (Fiore et al., 2008, CHI)
The skill itself runs as an 8-phase process: audit, discovery interview, reality check, photo strategy, copy, settings, implementation, algorithm strategy. The discovery interview is the core — about 40 structured questions designed to surface specific, hard-to-fake material, then that material gets written into prompts using the signaling framework.
It works well within Claude's constraints. The skill format turns out to be a good fit for this kind of structured, consultative process — the conversation architecture maps naturally to an interview that builds on itself.
MIT licensed. The full research document with citations and evidence tiers is in the repo.
https://github.com/b1rdmania/hinge-profile-optimizer
I wanted to build something that actually helps people fix their dating profiles. Not tips. A proper process. I used Claude to run parallel research agents across the academic literature on dating app behavior — signaling theory, self-presentation, deception linguistics, mate selection economics. The agents pulled in findings from about 45 papers, and I sorted them into three tiers: peer-reviewed, platform data (Hinge/OkCupid's own numbers), and conventional wisdom. The research doc in the repo is honest about which is which. Some findings that shaped the design:
Vague language is a linguistic marker of deception. People who lie in profiles use abstract, generic phrases. People who tell the truth use concrete, specific language. (Toma & Hancock, 2012, Journal of Communication) Desirability on dating apps follows a power law. The top profiles get 10-100x the attention of the median. You cannot out-volume this curve. (Bruch & Newman, 2018, Science Advances) "I have a great sense of humor" is what signaling theory calls a cheap signal — anyone can say it. Actually being funny in your prompt is a costly signal — you need the ability. Costly signals are processed as more credible. (Donath, 2007, building on Zahavi's handicap principle) Creative, original profile text independently predicted attractiveness ratings beyond photos. (Fiore et al., 2008, CHI)
The skill itself runs as an 8-phase process: audit, discovery interview, reality check, photo strategy, copy, settings, implementation, algorithm strategy. The discovery interview is the core — about 40 structured questions designed to surface specific, hard-to-fake material, then that material gets written into prompts using the signaling framework. It works well within Claude's constraints. The skill format turns out to be a good fit for this kind of structured, consultative process — the conversation architecture maps naturally to an interview that builds on itself. MIT licensed. The full research document with citations and evidence tiers is in the repo. https://github.com/b1rdmania/hinge-profile-optimizer