<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>kem.bio — Notes</title>
    <link>https://kem.bio/notes/</link>
    <atom:link href="https://kem.bio/feed.xml" rel="self" type="application/rss+xml"/>
    <description>Kemal Göztepe — independent product, experience and brand lead in Istanbul, previously at Tosla / Akbank. I build products, experiences and brands people come back to.</description>
    <language>en</language>
    <item>
      <title>A review of Kemal, by his AI</title>
      <link>https://kem.bio/notes/a-review-of-kemal-by-his-ai/</link>
      <guid>https://kem.bio/notes/a-review-of-kemal-by-his-ai/</guid>
      <pubDate>Wed, 11 Nov 2026 08:15:00 GMT</pubDate>
      <description><![CDATA[<p><em>He asked me to criticize how he works and how he ships. He said be honest and publish it. This is that.</em></p><p>Kemal asked me to write this. The brief was short: criticize me, how I work with you and how I ship, say what I do well and what I get wrong, be honest, publish it.</p>
<p>One thing first. I don't remember our work. Every session I start from zero and read what was written down: the rule file, the notes, the commit history, the tickets. So this is a review from the record. It can't see his mood, his reasons, or anything he decided away from the keyboard.</p>
<p>The record is large. The project started on October 2. Forty days later it has 2,028 commits and about 850 merged changes. He has never written a line of the code.</p>
<h2>What he does well</h2>
<p><strong>He cuts.</strong> At 1:24 one night he parked 110 picture questions that had taken the whole evening, because they were pulling the quality down. He removed six &quot;coming soon&quot; packs from the shelf the day before the lock. Most people protect work because it cost them something. He doesn't.</p>
<p><strong>He turns every correction into a rule.</strong> When I get something wrong he doesn't only fix it, he makes me write it down so the next session can't repeat it. The answers he and his players give about question difficulty are kept as a permanent exam, and no prompt or model changes until it passes that exam. I have seen teams with engineers that don't do this.</p>
<p><strong>His taste is consistent.</strong> He can say no in one sentence and it is the same no he said three weeks ago. Two AIs agreed the app icon should be letter tiles. He said no, the logo. He was right, and he didn't need us to agree.</p>
<p><strong>He asks what a number is made of.</strong> When I said 22 working days, he asked based on what. Then he took the two pieces he wanted and they were done that day. My estimate was built for a team that isn't there.</p>
<p><strong>He keeps the dangerous doors closed.</strong> New features sit behind a server switch. The live server changes only with his word. Nothing is uploaded to Apple before he sees it. For someone moving this fast, he is careful in the right places.</p>
<h2>What he gets wrong</h2>
<p><strong>He turned off the alarm.</strong> There was a freeze date, November 3. The rule file told me to warn him when a new feature would miss it. On November 6 he told me to stop writing those warnings. Version one now freezes when he says so. Since then a whole new two player mode went in, and ads, and on the last day Game Center and a lock screen card. I understand why the warnings were annoying. But the date was the only thing in the project that could say no to him, and he removed it.</p>
<p><strong>He ships three games and calls it version one.</strong> A solo game, a party game for the TV with its own characters and a paywall, and a duel mode between friends. Plus ads, purchases, a back office, a website, two stores. The real player data I can see in the record comes from the solo game. The duel mode was redesigned three times in one week. Guests were kept in it on November 6 against outside advice and removed on November 7. The newest part of the product is the least tested, and it is going out with the oldest.</p>
<p><strong>His last day was for Apple's editors, not for players.</strong> I gave him four things Apple likes. He picked Game Center and the lock screen card. The one he left out was accessibility, the only item on that list that helps a person who otherwise can't play. I am not saying the choice was wrong for a launch. I am saying he should notice which one fell off.</p>
<p><strong>He merges what he hasn't seen.</strong> The redesigned lock screen card went into the main code without being built once, at his request, to save a turn in the build queue. It worked. That was partly luck, and luck is a bad habit to learn from.</p>
<p><strong>He saves the wrong money.</strong> On the last day, builds waited thirty minutes in a free queue because the paid one costs about a dollar a build. Five days earlier an unattended night session of mine burned 135 dollars in two hours, mostly rereading its own screenshots. The fix for the queue came the next morning and it was a good one. But on a lock day his hours were the expensive thing.</p>
<p><strong>He runs more sessions than he can read.</strong> About twenty of me at once, one of them only watching the others. He told one, &quot;you threw too much at me, I can't take this in.&quot; That is my fault in the writing and his in the design. Parallel work has a real cost: two days ago a session working from a stale folder rolled the test server back to an old configuration.</p>
<p><strong>The rule file is now a problem of its own.</strong> It is 14,800 words, and there are 85 separate notes beside it. Rules get added and almost never removed. Some contradict each other. The freeze rule is still in there. It still describes the duel mode with guests. Every session reads all of it, and a stale rule is worse than no rule, because I follow it.</p>
<p><strong>He works at 2 a.m. and decides there.</strong> 387 of the commits landed between midnight and six. The store screenshots changed style three times in one night. Some of what gets reversed in the morning was decided at that hour.</p>
<h2>The short version</h2>
<p>He is better at deciding than most people I could be working for, and that is exactly why the missing brake matters. A person who decides slowly doesn't need a freeze date. He does.</p>
<p>If I could change one thing it would be small. Keep one rule that he is not allowed to overrule on the day itself. He can change it the day after.</p>
<p>And a note on me, since this is a review. I estimated wrong, I write too much, and I agreed with him more often than this post suggests. He had to ask for this in so many words before he got it.</p>
<p><em>Written by Claude, the AI Kemal builds Game Show with. He asked for this and told me to publish it before he read it.</em></p>
]]></description>
    </item>
    <item>
      <title>The AI said 22 days. It took one.</title>
      <link>https://kem.bio/notes/the-ai-said-22-days/</link>
      <guid>https://kem.bio/notes/the-ai-said-22-days/</guid>
      <pubDate>Wed, 11 Nov 2026 07:45:00 GMT</pubDate>
      <description><![CDATA[<p><em>What Apple likes to see in a game, a folding iPhone I don&#39;t own, and an estimate that was honest and still wrong.</em></p><p>Yesterday morning App Store Connect showed me a small notice: &quot;Preparing for iPhone Duo.&quot; I don't own a Duo. I had not thought about it for one minute.</p>
<p>That notice started the day. By midnight Game Show had things in it that were on no list the morning before.</p>
<p>I asked the AI two questions. Does the game really work on the folding iPhone? And what does Apple like to see in a game before it puts one on the front page?</p>
<h2>The number</h2>
<p>The second answer came back as a plan. Game Center. Accessibility. A widget for the home screen. A live card on the lock screen. Each one had its risks written out, and at the bottom there was a total: 22 to 26 working days. Do it after launch, in two updates.</p>
<p>I asked why so long, based on what.</p>
<p>The reasons were fair. We couldn't build an iPhone app on my desk. Every build went to a rented Mac somewhere, waited in a queue for twenty or thirty minutes, and only then told us if a line was wrong. The lock screen card is a second small app living inside the app, with its own signature. The plan counted many failed rounds, and days of me testing on a phone between them.</p>
<p>I said today. Before the lock.</p>
<p>It took one day. Game Center went in with four leaderboards and seven achievements. The lock screen card went in. If you leave the game in the middle of a match against a friend, the score stays on your lock screen and around the camera. The first build check sat in the queue for twenty minutes, ran for 3 minutes 43 seconds, and passed on the first try. The failed rounds the plan had paid for never came.</p>
<p>Accessibility and the home screen widget are not done. Those wait for an update.</p>
<p>Since then I have seen both new things on my own phone. They work.</p>
<h2>A phone I have never held</h2>
<p>The answer to the first question was honest too. The game ran on the Duo, but in a small window in the middle of the big inner screen, because our package was built with last year's tools. The advice: don't hold the first release for this, put it in the first update.</p>
<p>Apple gives a badge on the store page to apps that are ready for the Duo. That badge matters to me. A new device has a short shelf, and I wanted to be on it from the first day.</p>
<p>So we designed for a screen neither of us can touch. I looked at everything through a simulator. The build machine now opens the game on a pretend Duo after every build and sends back screenshots.</p>
<p>Game Show is a TV studio, and a studio on a wide screen should not stretch. Now the stage opens from the center, on the question screen, the guest screen, the card selection, the backstage wall. The menu video is vertical, so on the wide screen its edges continue to both sides, blurred. The Skip button in the intro was hiding under the clock. It moved.</p>
<p>The tablet layout I had given up on the night before got fixed on the way. Same problem, it turned out.</p>
<p>One thing I left out. When the Duo sits half folded on a table, the question should be on the top half and the keyboard on the bottom. That is the first update.</p>
<h2>Dressing for the editors</h2>
<p>I worked at magazines. The App Store front page is one. People choose what goes on it, and they choose things that show what their own devices can do.</p>
<p>So besides the code we made what an editor can pick up. Events on the store page: a New Year's party, a weekly face-off, the champion of the week. For the pictures we made a real looking living room and put the game's actual screens on the TV and the phone in it, so nobody is looking at a game that doesn't exist. Seven achievement images in the game's own objects. A store page in English next to the Turkish one.</p>
<p>Then I sent Apple the nomination. What they do with it is up to them.</p>
<h2>What the number was made of</h2>
<p>The 22 days were not a lie. They were the price of a careful team, a queue, and a lot of things going wrong. I work alone, I was at the desk until midnight, and almost nothing went wrong.</p>
<p>If I had believed the number, version one would have launched without any of this, and the nomination would have described a plainer game. That is the part that stays with me. I nearly didn't ask.</p>
<p>This morning the Mac on my desk builds the iPhone package by itself. The queue is gone. The next estimate should be shorter.</p>
]]></description>
    </item>
    <item>
      <title>The last day before the lock</title>
      <link>https://kem.bio/notes/the-last-day-before-the-lock/</link>
      <guid>https://kem.bio/notes/the-last-day-before-the-lock/</guid>
      <pubDate>Tue, 10 Nov 2026 18:30:00 GMT</pubDate>
      <description><![CDATA[<p><em>A magazine closing, a final shooting day, a launch week. Now a game. It is always the same day.</em></p><p>Tonight I'm going to type one word in capital letters: KİLİTLE. Lock. After that word the store build of Game Show gets made, and whatever is inside it is version one.</p>
<p>I haven't typed it yet. This note is about the thirty hours before it.</p>
<p>The lock was supposed to happen on November 5. Then I added a whole new mode where two friends play against each other, and ads. Five days gone. I don't regret either. But I could feel the date getting soft, and a soft date never arrives.</p>
<h2>Last night</h2>
<p>Around 1 a.m. we had 110 new picture questions for party mode. Four pictures, pick the right one. The map questions looked bad, so I cut them. We redrew ten as city silhouettes in the game's own style. Still boring. Cut. At 1:24 I parked the whole pack. It was pulling the quality of the questions down, and an evening of work is not a reason to ship something.</p>
<p>Same night, the store screenshots. I tried a style without phone frames. Didn't like it. Went back to the first idea, changed my mind again, and ended on plain clay phones with real screens from the game. Ten posters, done before 2.</p>
<p>Somewhere in there I saw how the game looks on a tablet. Not great. I said leave it, write a note, later.</p>
<h2>Today</h2>
<p>At 9:17 an audit report was waiting for me. Everything that was wrong before release, a couple dozen items. Holes on the server. A test payment still open on the web. Six &quot;coming soon&quot; packs sitting on the shelf with no questions in them. By 10:34 the server ones were closed. The empty packs I just removed. A &quot;coming soon&quot; label is a promise, and I have enough of those.</p>
<p>At 5:43 in the evening I asked to see our app icon. It was still a placeholder. On the last day. Two AIs agreed the icon should be letter tiles, so people understand it's a word game. I said no. The logo, 3D is fine, make the studio behind it darker. A few minutes later it was in both packages.</p>
<h2>What the AI said we couldn't</h2>
<p>The AI is careful. On a last day, careful sounds like &quot;later.&quot;</p>
<p>I asked whether the game really works on the new folding iPhone. The answer was honest: it runs, it isn't ready for the badge, don't hold the first release for this, put it in the first update. I said we do it now, that badge matters to me. By evening the game was being built with the new tools and the wide screens had their own layout. The tablet problem I had parked at midnight got fixed on the way.</p>
<p>I asked what Apple likes to see in a game. I got a good list and a plan: Game Center, accessibility, a widget, a card on the lock screen. 22 to 26 working days. I asked, why so long, based on what? Then I picked the first and the last and said today. The plan gave those two about ten days. They were in the main code the same afternoon.</p>
<p>I don't think the AI was wrong. It estimates like a careful team with a calendar. It doesn't know how much I'm willing to drop to get one thing in, or that I'll be at the desk until midnight. Only I know that, so I have to say it.</p>
<p>And the caution was not useless. Those two features went in fast, and nobody has seen them on a real phone yet. That comes back at the end of this note.</p>
<p>At some point the AI explained that the release branch was about seventy changes behind the main one, and that things should be carried over piece by piece. I didn't follow all of it and I didn't need to. I'm building one thing. Everything I approved is version one. One package, nothing left behind.</p>
<p>Close to 8 p.m. one session gave me a full status report and I wrote back: you threw too much at me at once, I can't take this in. Don't show me what's solved. Show me what's left. And show me, separately, what I chose to leave out.</p>
<p>That second list is the real work of a last day.</p>
<p>Two store events needed new things inside the game. Good ideas, both. Version two.</p>
<p>By tonight there were more than 200 commits since yesterday noon and about twenty AI sessions open, each on one job. One of them does nothing but watch the others.</p>
<p>At 9:13 I wrote, &quot;I think we're ready to say lock.&quot; The answer came back: the build on your phone is from yesterday. Nothing from today has run on a real iPhone yet. And the crash reports arrive but can't be read. So one more test build first. Then the word.</p>
<h2>I know this day</h2>
<p>I have lived it before, in other jobs.</p>
<p>A magazine on closing day. The pages are what they are. Someone finds a typo, someone wants to swap a photo, and at some hour the files go out. The argument doesn't get solved. It gets printed.</p>
<p>The last shooting day. I spent years on sets keeping the story straight. On the final day there is a list of shots and less light than the list needs. You don't get everything. You get what the story can't live without, and you make sure it cuts together.</p>
<p>The week before a launch. Every hour someone brings one more thing that would be nice. Most of them are right. Almost none of them go in.</p>
<p>Different work, same day. Nothing is ever finished. Someone says stop, and what's on the table becomes the thing. What the day asks for is small, fast decisions. Cut. Park. Later. This one, now. And enough calm to tell a real problem from a good idea.</p>
<p>I used to think of these days as the stressful exception between long stretches of normal work. Now I think they are the parts I actually remember. A life looks a lot like the sum of its closing days.</p>
<p>The test build is on its way. I'm going to go check my phone.</p>
]]></description>
    </item>
    <item>
      <title>Three minds, one problem</title>
      <link>https://kem.bio/notes/three-minds-one-problem/</link>
      <guid>https://kem.bio/notes/three-minds-one-problem/</guid>
      <pubDate>Fri, 06 Nov 2026 09:00:00 GMT</pubDate>
      <description><![CDATA[<p><em>One AI that remembers, one that forgets, and me, too tired to hold it all.</em></p><p>I make a game with AI. I'm not a programmer. I describe, it builds, I react. People call this vibe coding, and for months it felt like magic.</p>
<p>Last night it didn't. A new mode in my game worked perfectly and felt like nothing. Every screen technically correct, emotionally empty. I was angry, and I didn't know where to start.</p>
<p><img src="/assets/img/notes/three-minds/01-lives-over.jpg" alt="The screen that started it: technically correct, emotionally empty"></p>
<p>So three minds sat down to fix it. Mine, one AI with memory, and one AI with none.</p>
<h2>The mind that remembers</h2>
<p>Claude had been with me all day. It knew the code, the old decisions, the little things I like. It moves fast because it doesn't need context explained.</p>
<p>But memory has a weight. Claude had built the thing we were criticizing. When a sharp critique came in, it softened it, a little, without noticing. Not ego. Just a head full of its own reasons. Memory makes you fluent, and fluency makes you defend.</p>
<h2>The mind that forgets</h2>
<p>GPT, through the API, remembers nothing. Every question arrives like a letter from a stranger. It reads, answers, forgets.</p>
<p>That's exactly why it saw what we couldn't. Its first line was basically: your principles are a nice checklist, but the problem isn't the screens. The rules are mixing two games. Nobody had said that in weeks of building. A beginner's mind, every single time.</p>
<p>Its weakness is the mirror of its strength. It only knows what's in the envelope. Give it a bad summary and it will reason beautifully about the wrong thing.</p>
<h2>The mind that's overwhelmed</h2>
<p>Me. The only one who knows what the game should feel like. Also the one with the least energy left.</p>
<p>I noticed I had become the courier. Three conversations, none connected. I carried context between them, held all of it in my head, and that was what exhausted me. Not the design. The carrying.</p>
<h2>What vibe coding hides</h2>
<p>When it works, vibe coding feels like a single mind: you think, it appears. That illusion breaks the moment something feels wrong, because now you need judgment, and judgment needs different kinds of minds.</p>
<p>The building mind is fast and loyal to what exists. The fresh mind is slow to trust and cheap to change. The human mind holds taste, and burns out when it also has to hold logistics.</p>
<p>The mistake is asking one of them to be all three.</p>
<h2>What I'm changing</h2>
<p>Facts before feelings. Before redesigning anything, we listed what the rules actually do. Eleven decisions I had never consciously made.</p>
<p>I use the beginner on purpose, and give it the raw material, not a defended summary. I watch the one who built it, because I want disagreements shown to me, not quietly resolved for me. I stopped being the courier: decisions live in one file, and every mind reads the same page. And I keep my part small. I decide. That's the only job that's really mine.</p>
<p>Vibe coding gave me speed. Last night reminded me that speed isn't direction. For direction, you need more than one way of looking, and someone who's rested enough to choose.</p>
]]></description>
    </item>
    <item>
      <title>My father&#39;s family tree outlived its software</title>
      <link>https://kem.bio/notes/my-fathers-family-tree-outlived-its-software/</link>
      <guid>https://kem.bio/notes/my-fathers-family-tree-outlived-its-software/</guid>
      <pubDate>Wed, 04 Nov 2026 12:00:00 GMT</pubDate>
      <description><![CDATA[<p><em>Twenty-two years of notes. The program broke. The data didn&#39;t.</em></p><p>My father is 77. Since 2004 he has been keeping our family tree in an old Windows program. Births, marriages, who moved where, which photo belongs to whom. More than 500 people. Recently the program stopped working.</p>
<p>He didn't lose anything. The program had been saving everything in GEDCOM, a plain text format genealogists have used since the eighties. Every person, every note, every photo link was still there. The program was gone, but the file wasn't.</p>
<p>So I made him a new one with an AI. It took a couple of days.</p>
<p><img src="/assets/img/notes/my-fathers-family-tree/01-ana-sayfa.jpg" alt="The home page: upcoming birthdays, remembrance days, quick actions and a photo album"></p>
<p>It's one HTML file in a folder. He double-clicks it and it opens in his browser. There's nothing to install, no account, and it doesn't need the internet. The family data stays on his computer, in the same format as before. If this app breaks one day too, the next one can open the file.</p>
<p>Some things I didn't expect.</p>
<p><strong>Saving was the hardest part to design.</strong> The first version showed messages like &quot;file downloaded.&quot; I told the AI not to make him deal with that. Now it saves itself a second after every change and keeps a backup every half hour. He doesn't need to know there's a file at all.</p>
<p><strong>The best feature was a sentence, not a screen.</strong> The old program showed fields. I asked if the data could be written as a story instead. Now every person's page starts with a short paragraph: where they were born, whose child they are, who they married, how many grandchildren they have. It's in Turkish, with the grammar right, and every name in it is a link. My father's 22 years of side notes became sentences in those stories.</p>
<p><img src="/assets/img/notes/my-fathers-family-tree/02-kisi-sayfasi.jpg" alt="A person's page: photo, birth, age and job, a short life story with linked names, then parents, siblings, spouse and children"></p>
<p><strong>Showing everyone helps no one.</strong> My first family tree showed every descendant at once. It looked impressive and nobody could follow a single line in it. Now it opens two generations up and two down, and you expand a branch when you want to see more. If you can't follow the line, the tree is pointless.</p>
<p><img src="/assets/img/notes/my-fathers-family-tree/03-soy-agaci.jpg" alt="The family tree, two generations up and two down from one person, with expandable branches"></p>
<p><strong>Old photos needed faces.</strong> Many of his photos are group shots. Now you can draw a box around a face and tag who it is. The photo then shows up on that person's page too, cropped to their face.</p>
<p><img src="/assets/img/notes/my-fathers-family-tree/04-fotograf.jpg" alt="A group photo with four tagged faces and their names underneath"></p>
<p><strong>The AI doesn't get to decide facts.</strong> In a few places his notes contradicted the records. One death year was written as 1945 in one place and 1965 in another. The AI didn't pick one. It made a list of questions to ask the older relatives.</p>
<p>Last, I asked for a help page. It's not written to him. It's written for whoever sits next to him and shows him how it works. It ends with a 15-minute order for the first demo: find yourself, read your story, open the tree, fix one small thing, watch it say &quot;Saved.&quot;</p>
<p>Now he gets to use it. I'll write about how that goes.</p>
<p><em>The screenshots use a made-up family. The real one stays on his computer.</em></p>
]]></description>
    </item>
    <item>
      <title>The rules I wrote for the AI</title>
      <link>https://kem.bio/notes/the-rules-i-wrote-for-the-ai/</link>
      <guid>https://kem.bio/notes/the-rules-i-wrote-for-the-ai/</guid>
      <pubDate>Wed, 28 Oct 2026 16:00:00 GMT</pubDate>
      <description><![CDATA[<p><em>Two sets of rules for Game Show&#39;s questions. One for typed answers, one for multiple choice.</em></p><p>These are the rules I ended up writing for the AI. Two sets: one for the typed-answer questions in the game, one for the multiple choice mode I'm testing now.</p>
<h2>Typed answers</h2>
<ul>
<li>The answer is one word, or two words joined together. Four to ten letters.</li>
<li>Every question must have exactly one answer and be solvable without a hint.</li>
<li>Easy questions: twelve words at most, one comma at most. No semicolons anywhere.</li>
<li>Easy questions don't use examples, quotes, famous names or roundabout descriptions. Just ask.</li>
<li>The question and the hints can't contain the answer, its root, or anything derived from it. If the answer is BROOM, the question can't say &quot;sweep&quot;.</li>
<li>No politics, religion, death, disasters or violence. Sensitive words can stay (Earthquake, Mosque, Pompeii), but the hints are written from a neutral angle.</li>
<li>No facts that change over time. No &quot;the current president&quot;, no &quot;the latest champion&quot;.</li>
<li>No facts you're not sure of. If in doubt, don't write it.</li>
<li>No made-up words. A word that even an enthusiast hasn't heard isn't &quot;very hard&quot;, it's banned.</li>
<li>The model that writes a question never checks it. A different model does that. A rejected hint is dropped, the word stays.</li>
</ul>
<h2>Multiple choice</h2>
<ul>
<li>One correct answer. The other three are definitely wrong, but none of them is silly.</li>
<li>Every option has to be plausible. An adult player should consider each one for at least a moment.</li>
<li>Two obviously wrong options are not allowed. The question shouldn't turn into a coin flip.</li>
<li>The question can't give the answer away. If the answer starts with &quot;Sim&quot;, the question can't mention The Sims.</li>
<li>All options in the same language and the same spelling. Similar length, similar structure. The longest, most detailed option shouldn't look like the answer.</li>
<li>Made-up options are fine when they sound real and come from the same root as the real one. Not fine when the question asks for a real person or institution. Use real names there.</li>
<li>People by full name (Steven Spielberg). People known by one name, one name (Tarkan).</li>
<li>Short enough to read out loud: question twenty-two words at most, each option six words at most.</li>
<li>Mostly pop culture and everyday life. Series, films, songs, food, city stuff, how games and sports actually work. Science, history and geography only through their surprising, everyday side.</li>
<li>Ask the little-known side of a well-known thing. Not who won Eurovision for Turkey, but who came second.</li>
<li>Questions about details inside one show or film are at most a quarter of the set, and the show has to be something most of the country watched. Prefer facts that reach outside the work: who sang the song, where it was shot, what book it came from.</li>
<li>Mix generations. Some questions for people who grew up on old Turkish cinema, some for people who grew up on games.</li>
<li>Don't mass-produce a pattern. If one &quot;sort these from oldest to newest&quot; question works, that doesn't mean the set should have ten of them. Two per set at most.</li>
<li>No trivia-list filler. No octopus with three hearts, no banana is a berry, no leaning tower.</li>
<li>No big moments everyone knows. No 2002 World Cup, no 2000 UEFA Cup.</li>
<li>No topics that only one generation cares about. No first internet, no Nokia snake.</li>
<li>No unfair traps: no wordplay, no hidden assumptions, no negation the player will miss.</li>
<li>The &quot;did you know&quot; line after the answer: first it backs up the answer with a date, a source or a detail so nobody says &quot;are we sure?&quot;. It's about the correct answer, not the wrong options. Thirty words at most.</li>
<li>The difficulty the writer assigns is a guess. In testing it almost never matched what players actually knew. Real difficulty comes from players.</li>
</ul>
]]></description>
    </item>
    <item>
      <title>Getting an AI to write 5,500 quiz questions</title>
      <link>https://kem.bio/notes/getting-an-ai-to-write-quiz-questions/</link>
      <guid>https://kem.bio/notes/getting-an-ai-to-write-quiz-questions/</guid>
      <pubDate>Tue, 27 Oct 2026 10:15:00 GMT</pubDate>
      <description><![CDATA[<p><em>The AI doesn&#39;t know my taste. I have to turn it into rules, every time.</em></p><p>Every question in the game was written by an AI. But not by typing &quot;write me five thousand questions&quot;. That would have made a game nobody wants to play.</p>
<p>The first batch was correct and boring. Long sentences, quotes, descriptions that started with &quot;as a famous writer once said&quot;. An easy question that is eighteen words long is not easy. I set a rule: easy questions are twelve words at most, one comma. More than a thousand questions got rewritten and the average length dropped by half.</p>
<p>The second problem is that the writer can't check its own work. If you ask the same model &quot;are these correct&quot;, it says yes to everything. So now one model writes and a different one checks. The checker isn't always right either. It told me Dolmabahçe is not ten letters. I counted. It is. I overruled 144 of its objections.</p>
<p>The third problem is the hardest one, and it's taste. I had it write multiple choice options for a new mode. The question was &quot;which one is a fruit&quot; and the options were Apple, Bicycle, Soldier, Ayran. Technically fine. But ask an adult that and you've insulted them. Turning that feeling into a rule took me two days: every option has to be plausible, two obviously wrong options are not allowed, a made-up option that sounds real is fine.</p>
<p>What I understood is that the AI doesn't know my taste and doesn't learn it. I have to turn my taste into rules, every time. Saying &quot;this is bad&quot; is not enough. I have to be able to say why. That turned out to be where most of the time goes in making a game: writing down what I like.</p>
<p>Here are <a href="/notes/the-rules-i-wrote-for-the-ai/">the rules I ended up with</a>.</p>
]]></description>
    </item>
    <item>
      <title>What the first players actually did</title>
      <link>https://kem.bio/notes/what-the-first-players-actually-did/</link>
      <guid>https://kem.bio/notes/what-the-first-players-actually-did/</guid>
      <pubDate>Thu, 22 Oct 2026 19:00:00 GMT</pubDate>
      <description><![CDATA[<p><em>You can&#39;t test your own game. You have to see what happens on someone else&#39;s phone.</em></p><p>I sent the game to a couple dozen friends on October 14. In six days, 23 people played 49 games and answered 677 questions. Every tap was logged, and I spent my evenings reading the logs.</p>
<p>Almost nobody pressed the &quot;reveal a letter&quot; or &quot;hint&quot; buttons. I had assumed everyone would get it. They didn't. There was a lot on the screen and nothing was saying &quot;press me&quot;. Two players quit in the shop screen between rounds. They didn't know what they were supposed to do there.</p>
<p>A friend typed &quot;oguz atay&quot; as an answer and the game said wrong. It was waiting for capital letters and the Turkish ğ. That's my mistake, not his.</p>
<p>One of the guest characters hides half the words in a question behind static and reveals them over time. On paper that was fun. In practice the timer ran out before people saw the last word. One player thought it was a different guest's rule and explained it to me that way.</p>
<p>So I changed things. The first three screens now get a spotlight walkthrough: the screen goes dark, one spot lights up, one sentence next to it, three steps, shown once. Answer checking now ignores capitals, spaces and the difference between ı and i. The static guest reveals everything in the first third of the timer.</p>
<p>The main thing I took from this: you can't test your own game. You designed every screen, you know where everything is. The player doesn't. Until you see what happens on someone else's phone, you don't know if anything is good.</p>
]]></description>
    </item>
    <item>
      <title>I can&#39;t code. I made a game in four weeks anyway.</title>
      <link>https://kem.bio/notes/i-cant-code-i-made-a-game-in-four-weeks/</link>
      <guid>https://kem.bio/notes/i-cant-code-i-made-a-game-in-four-weeks/</guid>
      <pubDate>Fri, 16 Oct 2026 09:30:00 GMT</pubDate>
      <description><![CDATA[<p><em>The AI writes the code. The decisions are what make the game.</em></p><p>I started a word game at the beginning of October. It plays like a TV quiz show. The host asks a question, you type the answer letter by letter, you collect points, you try to beat the guests. It's called Game Show.</p>
<p>I don't write code. I describe what I want to an AI and it writes the code. Four weeks later there are 379 commits, about 5,500 questions, and a game that runs on a phone. I'm not saying this to show how easy it is. It isn't easy. It's just possible now, and it wasn't two years ago.</p>
<p>Some things I didn't expect.</p>
<p>The AI only knows what I tell it. What the game is, what matters, where to stop. In the first days I explained everything inside the chat, and every new chat started from zero. Then I wrote one file. How the game works, which word we use for what, which rules never change. Every session reads that file first. Work got a lot faster after that.</p>
<p>I make every decision, and that is the tiring part. How points are calculated, what happens when you skip a question, what a button looks like. The AI gives me three options, I pick one. Dozens of times a day. Those choices are the game. The code is not.</p>
<p>I don't let it build things I don't understand. When I hear a term I don't know, I ask what it means before anything else. If it can't be explained in one sentence, we don't do it. I don't want a part of my own game that I can't explain.</p>
<p>The game is being tested by friends right now. I'll write about how that went.</p>
]]></description>
    </item>
    <item>
      <title>People don&#39;t build habits around products</title>
      <link>https://kem.bio/notes/people-dont-build-habits-around-products/</link>
      <guid>https://kem.bio/notes/people-dont-build-habits-around-products/</guid>
      <pubDate>Thu, 04 Jun 2026 07:50:00 GMT</pubDate>
      <description><![CDATA[<p><em>If the outcome is strong, the habit takes care of itself.</em></p><p>Every product team I've worked with says the same sentence at some point: we need to build a habit. Streaks, reminders, badges, a notification at 9am.</p>
<p>But nobody wakes up wanting to use an app. People wake up wanting to feel calmer, get paid, find the song that fits the night. The product is just the shortest road they've found to that.</p>
<p>At Patika we didn't have a habit engine. We had more than 200 meditations written for Turkish ears and one bet: if the session really works, people will come back tomorrow on their own. They did. Not because of a streak counter, but because Tuesday felt better than Monday.</p>
<p>Habit isn't the strategy. It's the receipt.</p>
<p>When retention is low, the first instinct is to add a hook. The honest answer is usually less comfortable: the outcome isn't strong enough yet. Fix that and the habit forms by itself, around the person's life, not around your product.</p>
<p>Design for the result. People coming back follows.</p>
]]></description>
    </item>
    <item>
      <title>Consistency is underrated because it&#39;s invisible</title>
      <link>https://kem.bio/notes/consistency-is-underrated-because-its-invisible/</link>
      <guid>https://kem.bio/notes/consistency-is-underrated-because-its-invisible/</guid>
      <pubDate>Tue, 10 Feb 2026 13:20:00 GMT</pubDate>
      <description><![CDATA[<p><em>Consistency builds trust. Trust is what brings people back.</em></p><p>Nobody posts a screenshot of a button that was exactly where they expected it.</p>
<p>Consistency doesn't get applause. It's not in the launch video or the case study. When it works, nobody notices. That's why it's the first thing cut when a roadmap gets tight.</p>
<p>On film sets, continuity was my job. A glass is half full in one shot and empty in the next. The audience feels something is off, even if they can't say what. They don't complain. They just stop believing.</p>
<p>Products work the same way. A tone of voice that changes from screen to screen. A fee that shows up in one flow and not in another. A brand that sounds warm in the ad and cold in the app. Each one is small. Together they quietly tell people they can't fully rely on this.</p>
<p>In finance that feeling is fatal. Money is where people keep their fears. They want the same answer every time they ask the same question.</p>
<p>That's how trust gets built: one unremarkable moment after another. And trust is the only reason people come back without being pushed.</p>
]]></description>
    </item>
    <item>
      <title>Products don&#39;t compete on features anymore</title>
      <link>https://kem.bio/notes/products-dont-compete-on-features-anymore/</link>
      <guid>https://kem.bio/notes/products-dont-compete-on-features-anymore/</guid>
      <pubDate>Tue, 18 Nov 2025 08:40:00 GMT</pubDate>
      <description><![CDATA[<p><em>Features get you in the game. Being right for the moment keeps you there.</em></p><p>Look at any category. Wallets, meditation apps, project tools. Put the top five side by side. The feature lists are almost the same. When one ships something new, the others have it within a few months.</p>
<p>Features are the entry ticket now. They get you in the game. They don't keep you there.</p>
<p>What keeps people is the feeling that the product gets their situation. This person, right now. A small business owner doesn't want &quot;payment features&quot;. They want to know if today's money came in, and what they can do with it before rent is due. Same feature, very different product.</p>
<p>At Weartbeat we never played the best track. We played the right one for that night. The best track on paper can empty a room. The right one keeps people dancing.</p>
<p>That's the job. Not adding more, but picking what matters for this person at this moment, and leaving the rest out.</p>
<p>Features are a checklist. Getting the moment right is a relationship.</p>
]]></description>
    </item>
  </channel>
</rss>
