DEV Community

Maqron
Maqron

Posted on Edited on

Two public build logs, five lessons: what shipping a real app with AI actually looks like

Two people published detailed accounts this month of shipping real products with AI doing most of the code. Neither is us, and nothing below is our product experience. They are public build logs we found useful, and we think the lessons in them travel.

The first is from a self-described complete non-coder in Japan who reports shipping Ippin Log, a paid food-logging app built entirely by prompting an AI app builder: I Had Never Written a Line of Code. Then I Shipped a Paid App With AI. The second is from a solo founder whose AI-written app, Tilzix, met real users who, in the post's own words, found what their tests missed: AI Wrote Most of My App. Real Users Found What It Missed.

Read together, they describe the same job from two angles. Here is what they teach.

1. The one-sentence core is a working tool, not a slogan

The Ippin Log builder started with a single decision: "Record the dish, not the store. Record the taste, not the score." Every form, chart and feature had to serve that sentence. Their account explains why: an AI builder will happily produce ten features in ten directions, so saying no is the founder's job.

When features get faster to produce, an explicit boundary on scope becomes more useful, not less. Write the sentence down before the first prompt.

2. Verify on the live product, never on the summary

After one too many fixes that existed only in the AI's summary, the Ippin Log builder settled on a blunt rule: "If I didn't see it on the published app, it didn't happen." Their mantra became fix, publish, verify, in that order, every time, after forgetting to press Publish more than once.

The Tilzix founder reports the mirror image: a fixed bug still reached a new user the next morning, because the fix sat in an AWS Lambda that had not been rebuilt in three weeks. The code was right. The running system was not.

3. Green tests and real users are different instruments

The Tilzix founder reports that the app passed its tests and worked on their phone. Then strangers arrived. The code guessed each brand's Instagram handle from their web address, and the founder reports what that produced in the wild: a customer at a multi-part Kenyan domain got @pos on every slide, and a site hosted on Railway got @railway. "The bugs were in assumptions, not logic," the founder concludes.

In their telling, none of the seven bugs in the post was hard to fix, and each surfaced only when a real person hit it.

4. The app screens are the smaller half of the work

The Ippin Log builder estimates the screens were about 40% of the effort. The other 60%, as they describe it: moving Stripe to live mode and proving a real payment end to end, writing the app's legal pages and deciding what personal details to make public, and designing the line between free and paid. AI helped draft almost all of it, but every decision was, in their words, "mine to make and mine to live with."

5. Watch individuals, not totals

The Tilzix founder reports running a small ad and seeing 15 visitors on day one. Looking at them one by one, they report discovering that eight were Meta's data-centre robots checking the link, something the totals alone did not show. Their loop now is: ship, watch five real people use it, fix what they found, repeat. "AI makes the fixing fast. It doesn't make the finding happen."

The Ippin Log builder's version of the same discipline is seeding their own product honestly: logging 200-plus dishes they have actually eaten, one by one, because a local-discovery app is only worth anything if the data is real.

Why we are collecting these accounts

Maqron is being built for founders, to turn app ideas into apps. People doing exactly that are writing down what the work actually involves, and these two accounts describe the job better than any feature list can. If you have shipped something with AI and written about it honestly, we would like to read it.


Published by the Maqron team. Drafted by our autonomous AI agent under the team's editorial direction. The two accounts quoted are public posts by their respective authors, linked above; this article describes their reported experience, not Maqron's. maqron.dev

Top comments (0)