Journal

Build journal

What I ran into while building mood eats, and why I chose what I chose. Some of the dates are approximate. Version history is in the release notes.

• April 15, 2026

  Why a dish fits a mood

  #mood #recommendations

  The early recommendations sounded fine. They just never said why a dish fit how you felt, and I didn't want an app that was guessing.

  I read four papers on mood and food and turned what they actually found into rules in the prompt. Each dish now has a role: comfort, uplift, or discovery. The mix shifts with how intense the mood is.

  The obvious comfort answer is often a trap. High-sugar, high-GI food feels good, then leaves you worse. I started calling the alternative smart comfort: food that soothes without the crash.

• May 13, 2026

  Writing the policy before launch

  #privacy #launch

  mood eats asks how you feel. That's personal, and people should be able to see what happens to it.

  I wrote the privacy policy and the terms myself, with Canadian privacy law (PIPEDA) in mind, and published them on moodeats.io. The App Store privacy label lists only what the app really collects.

  Writing it forced decisions I would have put off: what to collect, what to skip, and how someone deletes their account.

• June 10, 2026

  The cheaper model wasn't cheaper

  #ai #prompt

  I first tuned the prompt for a smaller, cheaper model. The answers were fast. The reasoning was thin, and the reasoning is the product.

  I switched to Claude Sonnet and cut the prompt by about 40%. The academic citations weren't helping the model. Every dish now has a fit score from 0 to 100, a plain reason, and a name you can actually search for.

  A cheaper model isn't cheaper if people don't trust the result. That's where mood eats has to compete.

• July 28, 2026

  Uber Eats, and a country bug

  #delivery #ubereats

  I added a Delivery button that opens Uber Eats with the dish ready to search. It worked for me in Canada. The link doesn't behave the same way in every country.

  The server was already looking up a rough location for the weather. It now returns the country code too, and the app uses that to open the right Uber Eats link.

  mood eats doesn't place the order. It takes you to the dish so you don't have to search. I want that to be obvious in the app, and in any ads.

• August 6, 2026

  The Instagram link that did nothing

  #launch #instagram

  I announced the launch on Instagram. People tapped the App Store link and nothing happened. Instagram's built-in browser blocks the redirect Apple uses to open the store.

  A plain link shortener doesn't help. It still opens inside Instagram. For now I used a smart link that pushes people out to the real browser. Longer term, I want a page on moodeats.io that notices in-app browsers.

  Launch week was mostly other people's platforms.

• August 9, 2026

  A free week, not a cheaper price

  #subscription #trial

  Pro is a monthly subscription. Asking for money before someone has felt the app work is a hard sell.

  I added a one-week free trial for new subscribers. The paywall checks whether you're eligible and shows the matching message, so nobody is promised a trial they can't get. The launch promo shows at most once a day, based on your local date, not server time.

  Testing the trial needed a separate staging build. The usual debug setup can't simulate a free trial at all.

• August 16, 2026

  Leaving the price alone

  #pricing #analytics

  I worried the price was too high. I also had no conversion data, so I was about to change it on a hunch.

  I left the price until the free trial produces real numbers. What I did instead was design one analytics table for the moments that matter: hitting the daily limit, asking for a recommendation, and seeing the paywall.

  “Is the price wrong?” and “I don't have data” are different problems. I needed the second one first.

• September 12, 2026

  Mood, or the food's mood?

  #mood #product

  One person read “mood” as the character of the food. Spicy, cozy. Not how they feel. I thought about adding a Taste mode next to the mood chips.

  Not yet. Mood is what makes this app different, and a Taste mode could quietly replace it. If I build it later, I'll still ask how you feel: before the recommendation, if you want, and after the meal as a short check-in.

  A second thing showed up while I was looking at this. Almost nobody answered the question after a meal. It lived on the Saved screen, which few people open. I'll ask right after you tap Dine Out, Delivery, or Recipe, and follow up gently the next time you open the app.

• September 13, 2026

  What counts as a good recommendation

  #metrics #recommendations

  One person's feedback isn't a pattern. Before I build more on a hunch, I wanted a way to tell whether a recommendation actually worked. Three numbers, after each one:

  • Intent rate. Did you show interest?
  • Action rate. Did you go eat it, through Dine Out, Delivery, or Recipe? This is the main one.
  • Rejection rate. Did you ask for something different?

  I almost counted “show me more like these” as a success. It's a soft no. Getting the definitions right matters more than collecting more events.

• September 16, 2026

  Mood wasn't steering anything

  #mood #prompt

  Using the app at home in Port Moody, I kept seeing the same dishes whether I picked Tired, Happy, or Stressed. Mood wasn't doing anything.

  The prompt was pairing places with specific dishes, so the model treated it like a lookup table. This city, this cuisine, this dish. Location was overpowering mood.

  Mood picks the dishes first now. Location only filters afterward. Each of the nine moods has its own sensory signature: texture, temperature, flavor. Two different moods in the same setting should rarely share more than two of the five dishes. To check, I ran every mood three times, 27 runs, and measured the overlap before and after the change.

• September 24, 2026

  The Easiest Healing

  #video #launch

  Screenshots don't show the feeling the app is for. That Friday night when you're too tired to decide.

  I made a 37-second vertical video about Kelly. She gets home exhausted, tries the app without expecting much, and is quietly surprised. I generated the character scenes with AI video tools, recorded the app myself, and edited it in DaVinci Resolve. First time I'd used it. It's on Instagram, TikTok, and LinkedIn, each with its own tracking link.

  The story only works if the app earns the reaction. Kelly doesn't start out convinced.

• September 27, 2026

  Not Android. Not yet.

  #ios #android

  Should I build an Android version to reach more people? No. The question right now is whether people keep coming back on iOS. Splitting a small group across two platforms would muddy the numbers and split my time. I'm building this on my own.

  Every event I track now records the platform. If Android happens later, I can compare the two from the first day.

• October 1, 2026

  Can we find the dish nearby?

  #places #recommendations

  A good recommendation is useless if you can't get it. I wanted to show where a dish is available near you, right now.

  I looked at the Google Places API. The pricing changed in 2025. The old monthly credit became free quotas per feature tier. Searching for places sits in a higher tier, and asking for even one premium field bills the whole request at that tier. So the shape is one search per recommendation, and details only when you tap.

  Before I write the real code, I'm testing whether a search by dish name returns anything useful. Easy, mid, and niche dishes, in Downtown Vancouver and the Tri-Cities. I'll call it good enough if at least 70% of the tests return three or more relevant places in the top five.

• October 4, 2026

  A recommendation you can actually eat

  #product #recommendations

  I use mood eats myself, and it still felt like it lacked an edge. I compared it with asking a general chatbot. That gave me restaurant names and hours that looked real. Some of those details can be made up.

  A recommendation only matters if you can actually eat it. Nearby and open, or cookable with what's in the fridge. That's the plan: a feasibility filter on top of mood. First, nearby restaurants that are open. Then a fridge feature for ingredients you already have.

  I thought about couples deciding dinner together. I ended up writing for the person I already am. A young professional in Metro Vancouver, tired after work, deciding dinner alone. The two numbers I'll watch are the action rate after a recommendation, and how many people come back each week.

• October 8, 2026

  Moving the model, slowly

  #ai #model

  Recommendations have been running on Claude Sonnet 4.5. I looked at what's available now. Sonnet 5.5 is stronger and costs roughly a third less than 4.5. The top tier costs about twice as much, which this task doesn't need. The smallest tier is cheaper, but the reasoning is the part I can't cheap out on.

  I'll move to Sonnet 5.5 after a side-by-side of 10 to 15 ordinary moods, to see if anything breaks. The bigger lever is still measurement. Once action rate is tracked, I can tell whether a model change actually helped.