The AI keyboard for Hinge prompts and replies.

Hinge is a comment box wearing a dating app, and Pew's 21% figure is the demand signal

Pew Research Center's 2023 data show that 21% of U.S. adults ages 18–29 have used Hinge. That is smaller than Tinder's 42% in the same age band, which is the point. Hinge's audience self-selects for people who will read a prompt. The conversion event is not a swipe. It is a comment that proves you actually read the prompt, plus an early conversation that does not collapse into "so what do you do."

I built KOPY's Dating mode against that comment box. A Hinge like without a comment is a shrug. A comment that restates their prompt in different adjectives is also a shrug. The keyboard exists because I was losing matches to the blank composer while staring at "Two truths and a lie" like it was a take-home exam. The product constraint is the same iOS keyboard sandbox I hit on every app: Full Access for clipboard, ~70MB RAM, ephemeral URLSession, draft in under two seconds. The copy constraint is Hinge-specific. You are writing a public-ish note on their profile, then a private thread if they like you back.

Pew's easier-versus-harder split is why "thoughtful" cannot mean "long"

Under 30, Pew found 35% say apps make finding a long-term partner easier and 33% say harder. Hinge users skew toward the "easier if I try" camp, then punish you if the try looks automated. Length is not thoughtfulness. Specificity is thoughtfulness. A 12-word comment that names their Sunday ritual beats a 60-word paragraph about your attachment style.

When I generate on Hinge, I copy the prompt and their answer, not just "hey." The model needs the actual sentence. Apple's keyboard model does not let us scrape the Hinge UI. Clipboard is the API. If you copy nothing, you get a generic compliment and you deserve the silence that follows.

If you still need volume-first openers, that is Tinder first-message conversion. Stay on this page for prompt replies and the first five thoughtful turns.

Hinge KPI table: measure comments, not roses

Pew can tell you that 21% of 18–29-year-olds have used Hinge. Roses cannot. Roses are a scarcity mechanic, not a KPI. I do not publish invented KOPY usage stats. I publish the scoreboard I used while tuning prompt drafts. Count these for two weeks of sending.

KPI	How to count it	Hinge-specific target
Time-to-first-reply	Time from a like/comment you sent being accepted to your first in-thread message, or time from receiving a like to sending your first comment if you have not commented yet	Same waking-day. Hinge's "thoughtful" brand is not an excuse to wait 72 hours while you workshop a novel
Comment hit rate	Likes that become matches, divided by comments sent (exclude silent likes if you still send those; they pollute the denominator)	You want this higher than silent likes. If comments underperform silence, your comments are wallpaper
Continuation past 5 messages	Matched threads with five or more total bubbles after the comment	The prompt already did the icebreak. If you cannot get to five, you asked interview questions or you mirrored their joke without adding a second beat
Date-ask rate	Concrete plan asks divided by threads past 5 messages	Hinge daters tolerate an earlier ask if it is tied to the prompt ("museum Sunday" should become an actual museum window)

A secondary metric I like: prompt-noun reuse. In your first three messages, did you use a noun from their prompt that is not "coffee" or "travel"? If not, you are on Tinder energy inside a Hinge profile.

What shipping a keyboard taught me about Hinge prompts

Apple's custom keyboard docs force a design that looks clumsy until you use it: copy the prompt, switch keyboards, generate, paste. On Hinge that clumsiness is a feature. It slows you just enough to grab the right sentence. The failure mode I saw in TestFlight was people copying the last chat bubble ("haha") and generating a second "haha" with better grammar. Wrong clipboard, wrong job.

RAM is real. I cannot ship a huge on-device LLM in the extension. The ~70MB ceiling is why generation is a short network round trip with an ephemeral session. If I kept a background session, iOS could kill the keyboard or reject the architecture. The two-second budget matters on Hinge because people open the app in bed, see a good prompt, and will bounce if the assistant feels like a web form. The globe key has to feel like a tone picker, not a destination.

Dating mode tones I actually tap on Hinge: Funny when the prompt is already a joke, Rizz when the photo set is doing the work and the prompt is thin, Agree when they stated a preference I share and I can add a concrete example, Disagree when their take is spicy and they clearly want pushback. Spicy is rare on first comments. Hinge comments are visible enough that a sexual opener reads as a broadcast, not a vibe.

Prompt-reply patterns that survive the first filter

Pew's 30% of U.S. adults who have used a dating site or app, and 53% of 18–29-year-olds, means a lot of people have seen every "that's so real" comment already. I keep a short pattern list from my own sends:

• Callback plus fork: Quote one image from their prompt, then offer two next sentences they can pick. "Pretending you read philosophy on Sundays is elite. Camus in the gift shop, or you actually finished the book on the train?"
• Shared incompetence: If they admit a skill gap, match it with a specific one of yours, not "same." "I also cannot cook. I can, however, pick a taco place with unreasonable confidence."
• Tiny disagreement: If they plant a flag ("morning people are suspicious"), plant a smaller flag. "I will defend 9:30am. 5:30am is a cry for help."
• Plan seed: If the prompt is an activity, seed the date without asking yet. "That farmers-market photo is a threat. I need you to show me which stall is worth the line."

Patterns fail when I let the model keep the structure and drop the nouns. Then every comment becomes "Callback plus fork" with "coffee or drinks?" as the fork. That is how you recreate Tinder inside Hinge. For thread maintenance after the match, we wrote how to keep a Hinge conversation going with worked examples. This page is the research-and-KPI companion: what to measure, what the keyboard can and cannot see, and how Pew's usage numbers change the competitive set.

Early conversation: the five-message cliff on Hinge

Pew's 35/33 easier-versus-harder split under 30 shows up on Hinge as a warm comment that still dies by message four. Hinge matches often start warmer than Tinder matches because the comment already did work. That warmth creates a trap. People think the hard part is over. Then message three is "how's your week" and the thread becomes LinkedIn. Continuation past five messages on Hinge is a specificity stamina test.

My rule: each of my first five messages must contain either a proper noun from their profile, a timeword (Sunday, after work, this weekend), or a fork. If a draft from KOPY contains none of those, I regenerate with a different tone or I type the noun in myself. Insight mode is useful when they asked something that deserves a real answer (career change, moving cities). Social mode is useful when the chat has become logistical. Dating mode stays on until the flirt has a pulse.

Date-ask rate should be higher on Hinge than on Tinder for the same person, because prompts leak date ideas. If they wrote "typical Sunday = museum," and you wait until message twelve to suggest a museum, you wasted the prompt. Ask with a constraint: neighborhood or daypart. "Thursday after 6 near the park" beats "we should do that sometime."

Hinge keyboard sequence I use (distinct from Tinder batching)

Apple's custom keyboard model still requires Full Access for network generation, even when you only send five Hinge comments a night. I do not batch Hinge the way I batch Tinder. Hinge rewards fewer, better comments. I open the app once in the evening, pick five profiles I would actually date, and spend the clipboard-and-globe loop on those five. Full Access stays on because the keyboard still needs the network. I copy prompt plus answer as a single paste into the clipboard, sometimes with a slash between them so I remember both halves: "Typical Sunday = museum / pretending I read philosophy."

1. Select five profiles. Skip anyone whose prompts are empty or whose photos you cannot comment on honestly.
2. Copy prompt + answer. Switch to KOPY. Dating mode. Funny or Disagree first; Rizz only if the prompt is thin.
3. Generate in under two seconds. Add one noun the model missed. Send the comment. Do not send a second comment on the same profile.
4. When a match lands, copy their first in-thread message before generating. Do not reuse the prompt comment as the chat opener.
5. At message 6–8, ask for a plan that reuses a prompt noun. If they go dry, send one Continue-shaped fork, then stop.

Work mode almost never belongs on a first Hinge comment. It belongs if they are talking about a conference or a hiring loop and you are answering as a peer. Using Work mode as a personality is how you get "so what's your five-year plan" energy. I did that once. It read like a screening call. Never again.

Honest limits: what the extension cannot fix on Hinge

Pew can tell you 21% of 18–29-year-olds have used Hinge. It cannot make your photos decent. KOPY will not rewrite your prompts. If your own prompts are blank, you are asking other people to do work you refused to do. The keyboard drafts replies. It does not create a profile strategy.

I also will not claim the model "sounds like you" unless you edit. Voice lives in your nouns, your city, your dumb hobbies. The ephemeral request does not include your life story unless you put it in clipboard context or in the style settings we sync through the App Group. If you want the draft closer to you, put a one-line self-description in KOPY's style settings. Do not paste your therapy notes into a first comment.

Hinge is the thoughtful surface. Measure comments, five-message continuation, and date-asks that reuse prompt nouns. Use Tinder when you need opener volume. Use this keyboard when the blank comment box is the thing that used to cost you the match.