Daily Noodle · AI product
Shipping a full AI product as a team of one
Designing and building Daily Noodle solo, with AI: a journaling app where one question a day teaches an AI your emotional and behavioral patterns, then reflects them back with concrete things to try.
- AI product design
- Applied AI (LLM)
- Mental wellbeing
- AI-assisted build
- 0 → 1
Impact
A complete, live product, designed and built by one person
50 of 66 commits were paired with Claude (Sonnet 4.6, Opus 4.7 and 4.8). The build is genuinely AI-made, and version-controlled to prove it.
- Commits
- 66Commits to ship the live product
- AI-paired
- 76%Of commits paired with Claude
- Team
- 1Person, designing and building solo
- Stage
- 0 → 1From idea to a live, installable app
One question a day, and an AI that learns from it
Daily Noodle started as a strategic bet: that the retention problem above is a matter of product shape, not willpower. The shape I bet on is a single daily loop: one question that teaches an AI your patterns over time, then reflects them back with something to try.
Every day the loop repeats, and it compounds over a year.
The payoff arrives months after most people quit
Retention craters in weeks; the payoff only builds over months. The whole product exists to close that gap.
Retention craters in weeks; the payoff only builds over months. The whole product exists to close that gap.
Each AI tool owned a stage, and I owned the calls
The tools are swappable. What shipped a good product is the calls they don't make: which four mood families, what to cut, and when the convincing thing the model produced was wrong.
The tools are swappable. What shipped a good product is the calls they don't make: which four mood families, what to cut, and when the convincing thing the model produced was wrong.
Compressing 45 moods into four families you can read at a glance
Design started in Pencil.dev: fast exploration, with my judgment picking the direction. The clearest example is the mood system: ~45 moods are far too many to read on a calendar, so I compressed them into four families, ordered by intensity and colored for meaning.
Where judgment overruled the obvious choice
Muddy, hard to scan
Reads in half a second
Same year of entries, two groupings. Six categories muddied; four read in half a second, so I reverted. The git history still shows the arc.
I architected the product, and AI wrote the code
From design I moved into Claude Code and built the real thing: a typed React app, not a prototype. The split that matters for how I work: I architected the experience; AI executed the code.
And it ships continuously: Cloudflare deploys from the repo on every push, as an installable PWA, so every commit is a release and I iterate against the live product, not a local copy.
The product calls that don't show up in a screenshot
Privacy was the starting constraint
Journal entries are the most personal data a person has, so privacy was the starting constraint: a Postgres schema with row-level security, private by default. The AI features run as edge functions, so nobody's writing, and no API key, goes anywhere it doesn't have to. I owned the schema and security model; AI wrote the functions against it.
The Gemini key and your writing never reach the browser.
The Gemini key and your writing never reach the browser.
An AI that reflects your patterns back with something to try
This is the heart of the product, and mostly a design problem: a tool about your inner life can't feel like it's buffering or sound like a generic chatbot. Two Gemini functions do the work. One tags the mood behind each entry the instant you save; the other, diagrammed below, reads a whole year of them to find who you are.
It learns the patterns in how you feel and behave and the goals you keep circling, then ties them together: here's what's recurring, here's how it relates to who you're trying to become, here's one thing to try.
The constraints are the design: grounded in your own words, capped at four observations so it reads as clarity, not an audit, and responsible. When a heavy pattern persists, it points toward real help, not just another app screen.
A live product that turns a daily habit into self-understanding
The output of all of this is a complete, live product with one job: helping a person see themselves clearly.
The daily question removes the blank page. The resurfaced past answers let you watch yourself change. And the Reflect page assembles a year of writing into a streak, a mood heatmap, the themes you keep returning to, and the AI's observations. It's understanding you earn just for showing up.
Nobody fills out a goals form or drags a mood slider. The picture is a byproduct of the habit, which is the order of operations mental growth actually needs.
Responsive and installable: every surface reflows to a single column, and navigation drops to a thumb-reachable bottom bar.
Proven in daily use, with the retention test still ahead
Here's where I'll be straight: as a self-initiated project, this hasn't run long enough for a clean retention verdict, and I'd rather show you my thinking than invent a number. So I'll measure it against the three criteria I set at the start: what's validated, and what's still a hypothesis with a test attached.
People come back
I use it daily: the person most likely to quit (me) hasn't.
The patterns ring true
In my own use, the observations read as recognizably me. Needs a real n.
The insight moves something
Instrument whether suggestions get read, and taken.
Return rate vs. a blank-journal baseline
Day-7 and day-30 return: the number that proves the bet.
Would it matter at scale?
Back-of-envelope unit economics: at a category-typical ~4% free-to-paid, every 10K active users is ≈400 subscribers, on the order of $30K ARR per 10K, and it scales with reach. The point is the model, not the figure.
Execution got cheap. Judgment didn't.
I set out to build something genuinely personal, an AI that learns who you are and helps you grow, and to build it the way I now think product gets made: with AI as both the intelligence inside it and the team that ships it.
When a model can design the screen, write the code, and generate the insight, a designer's value doesn't disappear. It concentrates into the decisions: what to build, what to cut, how to make machine insight feel humane, and when the convincing thing the model produced is wrong. Here the delegate was AI; on a team it's engineers, PMs, and designers. The posture is the same: set direction, hold the bar. That's the version of this skillset I'd want a team to hire.
