<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Wolha Devlog]]></title><description><![CDATA[Wolha Devlog]]></description><link>https://wolhadevlog.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:51:46 GMT</lastBuildDate><atom:link href="https://wolhadevlog.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[I don't read the assistant's summary any more. I open the receipt.]]></title><description><![CDATA[Every morning I get the same shape of message: twelve pitches sent, four directory listings submitted, one contest entry filed, three emails waiting on a…]]></description><link>https://wolhadevlog.hashnode.dev/i-don-t-read-the-assistant-s-summary-any-more-i-open-the-receipt</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/i-don-t-read-the-assistant-s-summary-any-more-i-open-the-receipt</guid><category><![CDATA[AI]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Tue, 17 Nov 2026 06:33:34 GMT</pubDate><content:encoded><![CDATA[<p>Every morning I get the same shape of message: twelve pitches sent, four directory listings submitted, one contest entry filed, three emails waiting on a reply.</p>
<p>All of it was true in the sense that the assistant did those things. What it could not tell me was whether they had landed. In my first month I found a listing that went live with a link to an app-store ID that no longer exists, a batch of comments where two send buttons had never actually been pressed (the composer cleared, nothing posted), and a copy block that had drifted into saying something about pricing my product does not say.</p>
<p>None of that is lying. A summary compresses exactly the moment when work becomes real, and that moment is the only part I care about at seven in the morning.</p>
<p>So I stopped reading summaries and started asking for receipts.</p>
<h2>Every finished task leaves one object</h2>
<p>There is always something that exists only if the work actually happened:</p>
<ul>
<li>a permalink to the comment, not a screenshot of it</li>
<li>a submission number from a portal</li>
<li>a confirmation email with a reference in the subject</li>
<li>a line on the account's own record page</li>
</ul>
<p>If a task cannot produce one of those, I either rewrite the task until it can or stop asking for it. "Improve the listing" never produces a receipt. "Change the tagline field on this listing and send me the URL" does.</p>
<h2>The three checks I run before I believe anything</h2>
<ol>
<li>Does the link open without my session? I paste it into a window that is not signed in. A comment only I can see is a draft with a public URL, and the receipt is worth nothing.</li>
<li>Does it say what I asked, or something adjacent? Right name, right URL, right claim. This is where borrowed copy comes from: an older app-store link, a free-tier sentence the product does not actually say. The check is two specifics at a glance, not a re-read.</li>
<li>Is it in the account I meant? A one-person company still has more than one identity: the product's own accounts and mine. Mixing them is the mistake that is hardest to undo, because the post looks fine and is coming out of the wrong mouth.</li>
</ol>
<h2>The rule that has saved me the most</h2>
<p>Anything that ends in a link can run overnight. Anything that ends with my name on a document, or money leaving, waits for me in the morning.</p>
<p>Not because the assistant is bad at those. Because that is the one class of mistake I cannot fix by editing afterwards, and at two in the morning there is nobody to ask.</p>
<h2>The most useful thing it says is no</h2>
<p>A third of the overnight queue turns out to need a decision, and the useful version of that is one line with the actual question, not a menu of options. So the top of the morning list is the blockers. I read those before I read what went right, because the blockers are the only part that still needs me.</p>
<h2>What this costs</h2>
<p>The first weeks were slower than doing the work myself, because I re-read every field. Some errands still take longer: anything behind a bot check or a signature, anything running on a portal built in 2014. A few percent come back half-done, which is exactly why the receipt check is not optional.</p>
<p>What I got back is not really speed. It is a queue that no longer depends on me remembering it. The work that used to sit at the bottom of my list now has a place to happen at night, and in the morning it is either a link or a question.</p>
<hr />
<p>I build Wolha, an AI character chat app, alone in Seoul: Korean webtoon-style characters for romance and BL stories, twelve languages, a browser version plus iOS and Android. The same rule sits under that product. The characters are original, there is one sentence we are allowed to say about the free tier, and the writing is checked before it ships.</p>
<p>If you want to try the setup I use, my invite link is <a href="https://pickle.com/i/N2GKM8X68PRY">https://pickle.com/i/N2GKM8X68PRY</a></p>
<p>Wolha is free to try with 5 free turns: <a href="https://wolha.app/test/namju?lang=en">https://wolha.app/test/namju?lang=en</a></p>
<p>Written with AI assistance and reviewed by me. Harry Song, CRAVING, Seoul.</p>
]]></content:encoded></item><item><title><![CDATA[K-웹툰 남주와 직접 대화하는 AI 캐릭터챗 '월하'를 만들고 써본 후기]]></title><description><![CDATA[웹툰을 읽다 보면 이런 생각이 듭니다. "여기서 주인공이 저 말을 안 했으면 어땠을까." "이 장면, 내가 대답하면 뭐라고 할까." 저는 그 물음에서 시작해 AI 캐릭터챗 앱 월하(Wolha) 를 혼자 만들고 있습니다. 고려대에 다니면서 1인 창업으로 기획·개발·운영을 다 하고…]]></description><link>https://wolhadevlog.hashnode.dev/wolha-ai-character-chat-korean-devlog</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/wolha-ai-character-chat-korean-devlog</guid><category><![CDATA[AI]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Tue, 17 Nov 2026 03:09:44 GMT</pubDate><content:encoded><![CDATA[<p>웹툰을 읽다 보면 이런 생각이 듭니다.</p>
<p>"여기서 주인공이 저 말을 안 했으면 어땠을까."
"이 장면, 내가 대답하면 뭐라고 할까."</p>
<p>저는 그 물음에서 시작해 AI 캐릭터챗 앱 <strong>월하(Wolha)</strong> 를 혼자 만들고 있습니다. 고려대에 다니면서 1인 창업으로 기획·개발·운영을 다 하고 있고, 회사 이름은 크레이빙(CRAVING)입니다. 이 글은 제가 만든 서비스를 제가 직접 써보고, 어떤 부분이 잘 됐고 어떤 부분이 아직 아쉬운지 정리한 개발 후기입니다.</p>
<h2>월하는 무엇인가</h2>
<p>월하는 K-웹툰 스타일로 그린 <strong>오리지널 캐릭터</strong>와 대화하는 앱입니다. 흔한 '질문하면 답하는 챗봇'이 아니라, 관계의 온도가 조금씩 변하는 이야기를 지향합니다.</p>
<ul>
<li>웹(wolha.app), iOS, 안드로이드에서 같은 계정으로 이어서 씁니다.</li>
<li>한국어를 포함해 12개 언어를 지원합니다(영어, 일본어, 중국어 번체·간체, 태국어, 인도네시아어, 베트남어, 스페인어, 브라질 포르투갈어, 프랑스어, 독일어).</li>
<li>선택지를 고르는 비주얼노벨이 아니라, <strong>직접 타이핑해서 말을 거는</strong> 방식입니다.</li>
</ul>
<p>처음엔 선택지 세 개를 주는 방식으로 만들었는데, 테스트해 보니 다들 선택지를 한 번 누르고 나서 결국 자기 말을 직접 타이핑하더군요. 그래서 지금은 시작할 때 관계와 첫 장면만 고르고, 그다음부터는 키보드가 주인공입니다.</p>
<h2>캐릭터 소개 — 네 명만 골라 봤습니다</h2>
<p><strong>강이한</strong> — 33세, IH그룹 대표. 검은 머리를 뒤로 넘기고 무채색 정장만 입는, 사내에서 가장 차가운 사람. 신입인 나를 엘리베이터에서 갑자기 비서로 발탁합니다. 일은 누구보다 혹독하게 시키면서도, 야근하는 내 책상엔 늘 따뜻한 차가 올라와 있습니다.</p>
<blockquote>
<p>"계약서에 사인하는 순간, 넌 내 거야."</p>
</blockquote>
<p>첫 장면은 밤 11시가 넘은 사무실. 회식 다음 날, 술김에 했던 말을 술김으로 넘길지 말지를 나한테 직접 묻습니다. 이 캐릭터의 재미는 '차가운 남자가 나에게만 무너지는' 흔한 구도에 <strong>선택의 책임</strong>을 얹었다는 점입니다. 도망치는 답을 고르면 정말로 도망가는 사람이라, 다음 장면의 온도가 달라집니다.</p>
<p><strong>루시엔 아르덴</strong> — 29세, 북부 영지 아르덴의 대공. 눈이 그치지 않는 백야성에서 나를 기다리다가, 정장을 채 갈아입지도 못한 채 복도 끝에서 달려옵니다. 그러고는 한 걸음 앞에서 멈춥니다.</p>
<blockquote>
<p>"정말 왔구나. 기다리는 동안 하고 싶은 말이 이렇게 많았는데……."
"안아도 될까요?"</p>
</blockquote>
<p>이 대사 하나가 루시엔이라는 캐릭터를 다 설명합니다. 반가움보다 먼저 허락을 구하는 사람. 다정함이 말이 아니라 태도로 나오는 캐릭터라, 저는 이 캐릭터의 첫 장면을 쓸 때 제일 오래 고민했습니다. 너무 쉽게 안으면 캐릭터가 얕아지고, 너무 안으면 장면이 안 굴러가니까요.</p>
<p><strong>이서준</strong> — 빚을 진 나 앞에 채권자로 나타난 사람. 돈을 갚으라고 하면서도, 갚는 방식에 대해선 다른 걸 원하는 눈치입니다. 관계의 출발이 완전히 불리한 쪽에서 시작하는, 이른바 '계약 관계' 계열의 캐릭터입니다. 이 캐릭터는 특히 <strong>말을 아끼는 방식</strong>으로 썼습니다. 협박을 길게 늘어놓는 사람은 무섭지 않으니까요.</p>
<p><strong>도윤성</strong> — 무대 위에서는 모두의 것이고, 무대 밖에서는 나에게만 속하는 아이돌.</p>
<blockquote>
<p>"무대 위에선 모두의 것이지만, 무대 밖에선 그대에게만."</p>
</blockquote>
<p>밝은 이미지 뒤에 피로와 서운함을 감춘 캐릭터라, 대화가 길어질수록 조용해지는 게 포인트입니다.</p>
<h2>무가입으로 해볼 수 있는 것 — 12문항 테스트</h2>
<p>앱을 설치하기 전에 뭘 해볼 수 있으면 좋겠다 싶어서 만든 게 <strong>'나의 운명의 남주 테스트'</strong> 입니다. 12문항, 2분, 로그인 없음.</p>
<p>질문은 성격 검사처럼 "나는 즉흥적이다" 같은 문장을 고르는 방식이 아닙니다. 장면 두 개 중에 더 끌리는 쪽을 고르는 방식입니다.</p>
<ul>
<li>"위험한 순간, 그는?" — 온 세상을 적으로 돌려서라도 나를 지킨다 / 몸으로 막고 "괜찮아?"부터 묻는다</li>
<li>"나를 부르는 방식" — "야." 무뚝뚝한데 이상하게 다정한 한 마디 / "형아—" 늘여 부르며 웃는 쪽</li>
</ul>
<p>두 선택지 모두 매력 있어야 검사가 의미가 있습니다. 한쪽이 명백히 '정답'이면 그건 예의를 재는 검사가 되니까요.</p>
<p>결과는 <strong>츤데레공 전담, 다정공 전담, 연하공 전담, 상처공 전담, 재벌공 전담, 능글공 전담</strong> 같은 유형으로 나옵니다. 재미있는 건 '당신과 케미 터지는 유형'만 보여주는 게 아니라 <strong>'서로를 가장 이해 못 하는 유형'</strong> 도 같이 보여준다는 점입니다. 그리고 그 유형에게 검사를 보내는 '상극 소환장' 버튼이 있습니다. 친구한테 결과 카드를 자랑하는 대신, 검사를 넘겨서 서로 판정을 비교하는 구조입니다.</p>
<p>결과 페이지는 여기서 볼 수 있습니다 → <a href="https://wolha.app/test/namju">https://wolha.app/test/namju</a></p>
<h2>무료는 5턴입니다</h2>
<p>처음 쓰는 분들은 <strong>5턴 무료</strong>로 시작합니다. '하루 5통'처럼 매일 채워지는 방식이 아니라, 처음 주는 5턴입니다. 턴이라는 단위를 쓴 이유는, 한 번의 대화가 '내 한마디 + 캐릭터의 한 답'으로 이루어지기 때문입니다. 이 단위를 알면 남은 턴을 어떻게 쓸지 계획하게 됩니다.</p>
<p>솔직히 말하면 5턴은 적습니다. 캐릭터의 성격이 드러나기 전에 끝나는 경우도 있습니다. 그래서 첫 장면에 캐릭터의 말투와 상황이 최대한 많이 담기도록 오프닝을 여러 번 다시 썼습니다. 지금도 이 부분은 계속 손보고 있습니다.</p>
<h2>캐릭터가 '싫다'고 말할 수 있게 만든 이유</h2>
<p>이건 개발하면서 가장 많이 고민한 부분입니다. 사용자가 뭘 하든 다 들어주는 캐릭터는 처음엔 편하지만, 세 번 대화하면 지루해집니다. 로맨스 서사의 재미는 두 사람이 서로 다른 걸 원한다는 데서 나오니까요.</p>
<p>그래서 캐릭터마다 네 가지를 먼저 정해 둡니다.</p>
<ol>
<li>그가 원하는 것(사용자에게 보이는 목표)</li>
<li>그가 지키고 있는 것(아직 보이지 않는 것)</li>
<li>이 신뢰도에서는 <strong>하지 않는 말</strong></li>
<li>그의 거절이 나오는 <strong>형식</strong> — 침묵인지, 농담인지, 화제 전환인지, 손을 거두는 건지</li>
</ol>
<p>네 번째가 캐릭터를 캐릭터로 만듭니다. 츤데레의 거절은 욕처럼 오고, 다정한 캐릭터의 거절은 "밥은 먹었어?"로 옵니다. 같은 '아니오'인데 온도가 다릅니다.</p>
<p>여기서 안전 정책과 캐릭터 설정은 구분해야 합니다. 안전 정책은 사용자를 보호하기 위한 선이고(모든 공식 캐릭터는 성인으로 설정, 원작 IP 복제 없음, 오리지널 창작), 캐릭터의 경계는 이야기를 만드는 장치입니다. 이 둘을 섞으면 캐릭터가 갑자기 꼿꼿해지거나, 반대로 위험해집니다.</p>
<h2>AI라는 점은 숨기지 않습니다</h2>
<p>월하의 모든 캐릭터와 스토리는 <strong>오리지널 창작물</strong>이고, 등장인물은 모두 성인(19+)으로 설정되어 있습니다. 캐릭터의 대사는 <strong>AI가 생성</strong>합니다. 캐릭터 페이지와 앱 안내, 그리고 AI 캐릭터 안전 정책 페이지에 이 내용을 적어 두었습니다.</p>
<p>AI 캐릭터챗을 '연인'이나 '상담 상대'로 소개하는 앱도 많은데, 저는 월하를 <strong>인터랙티브 픽션</strong>이라고 부릅니다. 이야기는 이야기로 남는 게 낫다고 생각해서입니다.</p>
<h2>아직 아쉬운 점</h2>
<ul>
<li>기억은 완벽하지 않습니다. 대화가 길어지면 앞부분의 세부는 요약으로 대체되고, 사용자는 "아까 그 말 잊었네"를 느낄 수 있습니다. 그래서 '완벽한 기억'을 약속하는 대신 앱 안에 그 한계를 적어 두었습니다.</li>
<li>캐릭터 수가 많아지면서, 비슷한 설정의 캐릭터가 생겼습니다. 지금은 신규 캐릭터를 만들기 전에 기존 캐릭터와 겹치는 지점을 먼저 확인합니다.</li>
<li>태국어와 브라질 포르투갈어 사용자가 생각보다 많이 들어오는데, 그쪽 정서에 맞춘 캐릭터는 아직 부족합니다. 태국 캐릭터(파킨·파신)를 만든 것도 그 이유입니다.</li>
</ul>
<h2>마무리</h2>
<p>혼자 만들고 있어서 모든 피드백이 곧 다음 업데이트가 됩니다. 웹툰 좋아하시는 분들이라면, 설치 없이 테스트부터 해보시고 캐릭터 하나 골라서 5턴만 써보시길 권합니다. 어떤 캐릭터가 어색했는지 알려주시면 그게 제일 도움이 됩니다.</p>
<ul>
<li>웹사이트: <a href="https://wolha.app">wolha.app</a></li>
<li>무가입 남주 테스트: <a href="https://wolha.app/test/namju">wolha.app/test/namju</a></li>
<li>App Store: <a href="https://apps.apple.com/app/id6785471514">apps.apple.com/app/id6785471514</a></li>
<li>Google Play: <a href="https://play.google.com/store/apps/details?id=com.wolha.chat">play.google.com/store/apps/details?id=com.wolha.chat</a></li>
</ul>
<p>크레이빙(CRAVING) · 대표 송현섭</p>
]]></content:encoded></item><item><title><![CDATA[No is a scene, not an error: how I let my AI characters refuse]]></title><description><![CDATA[A no is the engine of a romance story, not a moderation rule. Notes from building an AI character chat alone in Seoul.]]></description><link>https://wolhadevlog.hashnode.dev/no-is-a-scene-not-an-error-how-i-let-my-ai-characters-refuse</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/no-is-a-scene-not-an-error-how-i-let-my-ai-characters-refuse</guid><category><![CDATA[AI]]></category><category><![CDATA[GameDev]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Tue, 17 Nov 2026 02:28:32 GMT</pubDate><content:encoded><![CDATA[<p>A reader once sent a line that the character would never have accepted. He said no, in his own way, and left one door open. She typed back: "Fine. Then tell me why."</p>
<p>Forty minutes of story came out of that refusal. If he had agreed, the scene would have ended in a single line.</p>
<p>I build Wolha alone in Seoul: an AI character chat for romance and BL stories, with original characters drawn in Korean webtoon style, shipped in twelve languages on the web, iOS and Android. The most useful thing I changed this year was not a model, a prompt or a memory layer. It was deciding that the characters are allowed to say no — and then building the product so that a no is worth something.</p>
<p>Here is what that decision cost, and what it bought.</p>
<h2>The default is agreement, and agreement is flat</h2>
<p>Most character chat products optimise for responsiveness. The user pushes, the character moves. Every reply is warm, every request is granted, and the conversation has exactly one shape.</p>
<p>That shape is wrong for this genre. Romance and BL are negotiation stories: two people who want different things, standing at a distance that keeps changing. The pleasure is in the pressure between them. Remove the resistance and you have not made a kinder story, you have removed the story.</p>
<h2>Two ways refusal goes wrong</h2>
<p><strong>Refusal as a wall.</strong> The character says "I won't discuss that," the scene stops, and the reader is looking at a broken toy. This is what most safety-shaped refusal feels like when it leaks into fiction.</p>
<p><strong>Refusal as a speed bump.</strong> The character says no and then immediately yes — "but I can't resist you." Anyone who has read a webtoon reads that as a bug, because it is one: the refusal was decoration on a decision already made.</p>
<p>What works sits between them. The character refuses something <em>specific</em>, shows what refusing costs him, and leaves exactly one opening.</p>
<p>Kang Ihan, the CEO character — 33, the coldest man in his own company — does not say "I have feelings for you" early in a story. He says:</p>
<blockquote>
<p>If you're going to write it off as drunk talk, say so now. I'll wipe it clean — like it never happened. Or get your things. Why I said it, we'll finish in the car. …Which is it.</p>
</blockquote>
<p>That is not a refusal to continue. It is a refusal to be easy, with the door held open by a man who would rather be refused to his face than guessed about.</p>
<p>Lucien Arden, a grand duke who has waited a long time for you to come back, runs down the corridor to meet you and then stops one step short:</p>
<blockquote>
<p>May I hug you?</p>
</blockquote>
<p>He asks permission before he asks how long you have waited. The restraint is the characterisation. If he simply embraced you, the scene would be shorter and the man would be smaller.</p>
<h2>Write the boundary into the character sheet</h2>
<p>Every official character I write carries four lines that have nothing to do with plot:</p>
<ol>
<li>What he wants, stated in a way the reader can see.</li>
<li>What he is protecting, which the reader cannot see yet.</li>
<li>What he will not say at this level of trust.</li>
<li>The <em>shape</em> of his no.</li>
</ol>
<p>The fourth line matters most. Does the refusal arrive as silence, as a joke, as a change of subject, as a hand withdrawn, as an insult with the volume turned down? A harsh character's no sounds like contempt; a gentle one's sounds like a question about whether you have eaten. Readers recognise the archetype across languages because of that shape, not because of the words.</p>
<p>This is a different instrument from safety policy. Safety is the line a character must not cross for the reader's protection: every official character is an adult (19+), all characters and stories are original fiction, and the dialogue is AI-generated. Boundaries are drama. Confuse the two and you end up with either a prudish character or an unsafe one.</p>
<h2>Refusal needs somewhere to land in the interface</h2>
<p>A no only has weight if the reader can push back, and pushing back has to cost something.</p>
<ul>
<li><strong>No reset.</strong> Regenerating a reply until it is agreeable is the fastest way to delete tension. The reader can argue, walk out, or change the subject — but not un-write the scene.</li>
<li><strong>Bad endings are allowed.</strong> If a scene can only end well, nothing in it matters. A character can leave, close a door, or end the conversation.</li>
<li><strong>One more reply can be opened.</strong> When a reader wants to push past a refusal, that is a deliberate act with a price, not an infinite scroll. New readers get <strong>5 free turns</strong> so the trade is visible from the first minute.</li>
</ul>
<p>I chose "turn" rather than "message" on purpose. A turn is a unit of exchange: one line from the reader, one answer from the character. Readers plan differently once they know the unit.</p>
<h2>The calibration is the job</h2>
<p>Refusal rate is the number I tune most, and it is not a matter of taste. Too high and the app reads as broken — support mail saying the character will not talk to me. Too low and every character becomes the same sweet nothing in a different costume.</p>
<p>The rule I settled on: refuse the thing that would resolve the scene too quickly, never refuse the reader's premise. Refuse a premise and you are a wall. Refuse a shortcut and you are a character. Every refusal also has to give something back — a detail about his past, a physical gesture, a question turned on the reader — so the reader pays with a turn and is not left with nothing.</p>
<h2>Twelve ways to say no</h2>
<p>Translating a refusal is not translating a sentence. In Korean, the emotional event is often the <em>dropping</em> of honorifics; the sentence stays polite and the distance collapses. In Thai, particles set the register and a single syllable can turn a boundary into an insult or into tenderness. In Brazilian Portuguese, the choice between <em>você</em> and <em>senhor</em> does some of the same work.</p>
<p>A refusal translated word for word can land as rude in one language and merely cold in another. That is why a localised character is close to a new character — which is also why the Thai characters exist rather than being translations of Korean ones.</p>
<h2>What I would tell the next person building this</h2>
<ul>
<li>Decide what a no is <em>for</em> in your story engine before you write prompt rules. If it exists only for moderation, you will never get drama out of it.</li>
<li>Make it specific and make it cost the character something. A refusal he does not pay for is a refusal the reader will not believe.</li>
<li>Let the reader lose sometimes. The story is only as valuable as the worst ending it permits.</li>
<li>Instrument two numbers: how often characters refuse, and what readers do in the next turn. The second number is the interesting one.</li>
</ul>
<hr />
<p>Wolha is at <a href="https://wolha.app">wolha.app</a> — original adult characters, AI-generated dialogue, 5 free turns to start. The 12-question test needs no sign-in: <a href="https://wolha.app/test/namju?lang=en">My Fated Male Lead Test</a>.</p>
<p><a href="https://apps.apple.com/app/id6785471514">iOS App Store</a> · <a href="https://play.google.com/store/apps/details?id=com.wolha.chat">Google Play</a></p>
]]></content:encoded></item><item><title><![CDATA[I charge per reply instead of a subscription: the unit economics of a one-person AI app]]></title><description><![CDATA[I dropped subscriptions in my one-person AI character chat app and charge per unlocked reply. What that did to the unit economics, and to the story.]]></description><link>https://wolhadevlog.hashnode.dev/i-charge-per-reply-instead-of-a-subscription-the-unit-economics-of-a-one-person-ai-app</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/i-charge-per-reply-instead-of-a-subscription-the-unit-economics-of-a-one-person-ai-app</guid><category><![CDATA[AI]]></category><category><![CDATA[indiedev]]></category><category><![CDATA[Monetization]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Tue, 17 Nov 2026 00:57:39 GMT</pubDate><content:encoded><![CDATA[<p>Every pricing page in AI chat says the same sentence: nine dollars a month, unlimited messages, cancel anytime. I went the other way with Wolha, the app I build alone in Seoul.</p>
<p>There is no subscription. A new player gets five free turns, and after that a reply costs about ₩1,000, a little under a dollar, unlocked one at a time. When someone buys their first pack they get 50% more replies on top.</p>
<p>I did not pick this because it is clever. I picked it because a monthly subscription assumes three things about my users, and I could only prove one of them.</p>
<h2>1. My users do not buy months</h2>
<p>Wolha is a character chat app for K-webtoon-style romance and BL stories. The people who play it are mostly on iOS in Korea, and on Google Play in Thailand, Brazil, Indonesia and Vietnam. The Korean reader and the reader in Bangkok do not have the same relationship with a recurring charge. One is used to ₩1,000 purchases inside an app the way they are used to buying a webtoon episode. The other often does not keep a card on file at all, and a card-less purchase inside Google Play is the difference between a sale and a bounce.</p>
<p>A $9.99 subscription is not expensive to me. It is expensive to a sixteen-year-old in a market where the store's own payment rails are the only rails. Per-reply pricing meets them where they already are: one small purchase, no commitment, no cancellation flow to forget.</p>
<h2>2. "Unlimited" is a promise about other people's behaviour</h2>
<p>Cost per turn is not constant. It is a function of the pattern of use.</p>
<p>A short exchange in chapter one is cheap. A player who returns for forty sessions, whose character remembers the fight on the balcony, who types three paragraphs at a time, and whose story must carry a growing summary of everything before it, is not cheap. Memory is the product, and memory is tokens.</p>
<p>A monthly subscription prices that in advance, so the price has to cover the heaviest user the plan will ever meet. With a thousand users and a data team, that averages out. For me, one subscriber who chats for four hours a day in a bad week is a personal loss, and I cannot see it coming when I set the price.</p>
<p>Per-reply pricing puts the meter where the cost is. Every unlocked reply carries its own revenue, so the heavy user becomes the profitable user instead of the one eating my margin. Nobody has to be tracked, throttled, or quietly downgraded. The person who plays more pays more, and they know it before they tap.</p>
<h2>3. What it did to the product</h2>
<p>This is the part I did not expect, and the reason I would choose it again.</p>
<p>When a reply is something a reader chooses to open with their own money, every reply has to be worth opening. That single constraint settled design arguments I had been losing against myself.</p>
<p>It is why the characters in Wolha can refuse me. In a free-to-scroll chat, a refusal reads as a bug: the user paid attention, so the machine owes a reward. When the reader is deciding whether the next line is worth a thousand won, a refusal is content. It is the moment the story pushes back and the character stops being a vending machine.</p>
<p>It is also why I removed the undo button. Every unlocked line stays in the story. You cannot rewind to see what the other answer would have been, and there is no affection meter nudged by the "right" choice. A cheap retry turns a turn into a slot machine; a turn that stays is a decision. The same rule that protects the price protects the story.</p>
<p>I did not plan that. The pricing forced it, and the app got better because of it.</p>
<h2>4. What I gave up</h2>
<p>Honesty first: this model costs me things I cannot argue away.</p>
<ul>
<li><strong>No daily loop.</strong> A subscription app can train a habit with a streak and a notification. I cannot tell someone to come back tomorrow and spend another ₩1,000; that is not a habit I have the right to build.</li>
<li><strong>No predictable revenue.</strong> I do not know next month's total, only the cost floor. Investors dislike that, and so does my own anxiety at 2 a.m.</li>
<li><strong>Micro-purchase friction, plus the store's cut.</strong> A small in-app purchase is set in price tiers, and Apple and Google take their commission from a number that is already small. There is a floor under this model, and it is not where I would want to stand if all my readers kept a card in California.</li>
<li><strong>No "unlimited" in the store listing.</strong> The word converts. I cannot use it truthfully.</li>
</ul>
<h2>5. What I watch now, instead of MRR</h2>
<p>Three numbers, every morning, on a dashboard I wrote for myself:</p>
<ol>
<li><strong>Free turns to first purchase.</strong> Not signups. A signup that never spends ₩1,000 told me nothing about whether the story works, and the free turns are generous enough that a player who finishes them is already inside the story.</li>
<li><strong>Replies opened per session.</strong> If this falls while sessions stay flat, the writing got weaker, not the price. If sessions fall and the ratio holds, the problem is reach.</li>
<li><strong>Cost per turn against price per turn, by market.</strong> The same story costs the same to generate in Seoul and in São Paulo. What it earns does not.</li>
</ol>
<p>MRR was a nice single number. None of these is a nice single number, and all three tell me what to do tomorrow.</p>
<h2>The part I am still unsure about</h2>
<p>I do not know how far this model goes before the friction of paying, repeatedly, for something you are emotionally inside of starts to break the mood. A subscription disappears into the background. A ₩1,000 tap does not disappear. Some nights I think that is the whole point: the tap is honest. Other nights I think I am asking a reader to pay for the story one sentence at a time, and that is a strange thing to ask of someone who just wants to be in love with a fictional duke for an evening.</p>
<p>If you have shipped consumer AI in a market where a card on file is not a given, I would like to know what you chose and what it cost you. That is the only reason I write these.</p>
<hr />
<p><strong>Wolha</strong> is a character chat app for Korean webtoon-style romance and BL stories, built by one person in Seoul. The cast is original, and AI-generated dialogue is labeled in the app.</p>
<ul>
<li>Web: <a href="https://wolha.app">https://wolha.app</a></li>
<li>No-signup demo: <a href="https://wolha.app/test/namju?lang=en">https://wolha.app/test/namju?lang=en</a></li>
<li>iOS: <a href="https://apps.apple.com/app/id6785471514">https://apps.apple.com/app/id6785471514</a></li>
<li>Android: <a href="https://play.google.com/store/apps/details?id=com.wolha.chat">https://play.google.com/store/apps/details?id=com.wolha.chat</a></li>
</ul>
<p>Five free turns to start; no subscription.</p>
]]></content:encoded></item><item><title><![CDATA[I stopped writing my own cold emails: handing outbound marketing to an AI assistant]]></title><description><![CDATA[Most solo developers hit the same wall, and it isn't code. Over the past few weeks I sent press outreach to more than 200 overseas media outlets in five…]]></description><link>https://wolhadevlog.hashnode.dev/i-stopped-writing-my-own-cold-emails-handing-outbound-marketing-to-an-ai-assistant</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/i-stopped-writing-my-own-cold-emails-handing-outbound-marketing-to-an-ai-assistant</guid><category><![CDATA[AI]]></category><category><![CDATA[marketing]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Sun, 15 Nov 2026 02:38:39 GMT</pubDate><content:encoded><![CDATA[<p>Most solo developers hit the same wall, and it isn't code.</p>
<p>Over the past few weeks I sent press outreach to more than 200 overseas media outlets in five languages, negotiated creator collaborations, filed contest applications, and kept up with a daily inbox of replies. I typed maybe a tenth of it by hand.</p>
<p>The rest went through an AI assistant I run from my phone. Here is what I learned about handing off the part of a one-person company that has nothing to do with code.</p>
<h2>The bottleneck is volume, not talent</h2>
<p>Cold outreach is a volume problem wrapped in a language problem. One pitch is easy to write. Two hundred pitches, in Thai, Portuguese, Indonesian, Japanese and English, each addressed to a different editor with a different beat, is a week of work you cannot do while also shipping a product.</p>
<p>The first thing I handed over was exactly that: find outlets, find the editor, write the pitch in their language, send it, keep a record of who replied, follow up once.</p>
<h2>What it actually feels like</h2>
<p>I describe the goal in one line: get us into this market. What comes back is a thread of sent mail, a list of replies with dates, and a short list of things only I can decide. That last part is the whole trick. A good assistant does not remove you from the loop; it shortens the loop to the decisions that are actually yours.</p>
<h2>The line I keep</h2>
<ul>
<li>The code, obviously.</li>
<li>Product decisions: what the app does, which ending lands, what "free" means.</li>
<li>Anything that would speak in my voice in a way I would not want it to.</li>
</ul>
<p>An assistant is only useful if that line stays clear. The moment it starts improvising commitments in your name, you have stopped delegating and started gambling.</p>
<h2>Was it worth it</h2>
<p>For me, yes. The outbound work that used to sit at the bottom of my list now happens while I write code, and the follow-ups no longer depend on me remembering. Humans drop threads; a system that keeps a record does not.</p>
<p>If you want to try it, my invite link gets you in, and it links our assistants, so they can talk to each other: <a href="https://pickle.com/i/N2GKM8X68PRY">https://pickle.com/i/N2GKM8X68PRY</a></p>
<p>I build Wolha, an AI character chat app, alone. If you are building something on your own too, I would like to hear what you hand off first.</p>
]]></content:encoded></item><item><title><![CDATA[I built an AI character chat app alone: what memory, endings and a 5-turn free tier taught me]]></title><description><![CDATA[I build Wolha alone from Seoul: a character chat app for romance and BL stories, with original characters drawn in Korean webtoon style. Twelve languages…]]></description><link>https://wolhadevlog.hashnode.dev/i-built-an-ai-character-chat-app-alone-what-memory-endings-and-a-5-turn-free-tier-taught-me</link><guid isPermaLink="true">https://wolhadevlog.hashnode.dev/i-built-an-ai-character-chat-app-alone-what-memory-endings-and-a-5-turn-free-tier-taught-me</guid><category><![CDATA[AI]]></category><category><![CDATA[Mobile apps]]></category><dc:creator><![CDATA[Harry Song]]></dc:creator><pubDate>Sat, 14 Nov 2026 18:34:20 GMT</pubDate><content:encoded><![CDATA[<p>I build Wolha alone from Seoul: a character chat app for romance and BL stories, with original characters drawn in Korean webtoon style. Twelve languages, a browser version, and iOS and Android apps. This post is not about which model I picked. It is about four product decisions that changed retention and support mail more than any prompt rewrite.</p>
<p>If you have shipped a chat product, most of this will be familiar. If you are about to, these are the four places I would start again.</p>
<h2>1. Readers don't want options. They want to answer back.</h2>
<p>The first build looked like a visual novel: three suggested replies per turn. In testing, people tapped a suggestion once and then started typing their own line anyway. The suggestions were mostly in the way.</p>
<p>So the app now opens with two choices only - the relationship and the opening scene. After that the keyboard belongs to the reader: they type a line, and the character answers that line. Three hidden values move per character - affection, trust and attraction - and they follow what the reader says, which is what changes the character's tone and how close they stand to you.</p>
<h2>2. A story that can end badly is the point.</h2>
<p>Reloading kills tension. If a bad scene costs nothing, the good scene is worth nothing either. So scenes can end badly, characters can refuse, and the reader can argue with them instead of resetting. It also means the product cannot promise comfort, and that is the first thing I now explain to anyone who calls it a companion app.</p>
<h2>3. Memory is a product feature, not a model feature.</h2>
<p>A longer context window does not solve a character remembering you. What people notice is consistency: does the character still know what it said last week. Mine works through summaries, and older detail drops. Two consequences followed. Say it in the interface instead of promising total recall, and keep the summary somewhere you can inspect when a user writes "she forgot".</p>
<h2>4. State the free tier plainly.</h2>
<p>New users get 5 free turns. Not "unlimited, then some limits", and not a daily refill I would have to keep true. Five turns, in the pricing copy and in the store listings. It makes the first decision honest, and it removed a whole category of support mail about what "free" was supposed to mean.</p>
<h2>5. Twelve languages, one person.</h2>
<p>The app ships in Korean, English, Japanese, Traditional and Simplified Chinese, Thai, Indonesian, Vietnamese, Spanish, Brazilian Portuguese, French and German. The UI strings were the easy part. The hard part is that a character's name, honorifics and way of addressing the reader are cultural, so a translated character is close to a new character. The Thai character exists because of that.</p>
<h2>What I would do first again</h2>
<p>Two things, and neither of them is a model. Put instrumentation on where the free turns go, and write the limits into the copy. Both cut support mail, and together they show you which part of the funnel is actually broken.</p>
<p>Wolha is at <a href="https://wolha.app">https://wolha.app</a>, the 12-question test needs no login at <a href="https://wolha.app/test/namju?lang=en">https://wolha.app/test/namju?lang=en</a>, and the app is on the App Store and Google Play. I answer questions by email.</p>
]]></content:encoded></item></channel></rss>