From Vibe Coding to AI Agent: A Complete Journey of Building a Baby Feeding System

This year has been momentous for me: my baby was born. Amid the joy of new life, there’s no shortage of anxiety—worrying about feeding the baby scientifically and whether something might slip through. At the postpartum care center, the confinement nanny kept regular feeding records—when the baby drank milk, how long they slept, how many bowel movements, whether they took AD supplements, etc.—all tracked to monitor feeding status.

But these were recorded with pen and paper, not digitized.

Once home and caring for the baby ourselves, I still felt it was necessary to record the baby’s care status: first, to track daily feeding details for reference; second, I wanted to analyze the baby’s condition based on daily feeding data and provide targeted advice—avoiding overfeeding or underfeeding—and also track growth trends.

The idea was there, so I got to work.

Thanks to a year of AI development, AI Coding has crossed the usability threshold and can now genuinely participate in real project development. So I planned to use AI to build a system that records the baby’s feeding and growth status—ideally without me writing any code myself.

Additionally, with the advancement of AI Agents, I started thinking about how AI could step into daily baby feeding as a dedicated AI feeding assistant. This article will cover my practices.

Requirements Analysis

Although a feeding system is a niche and highly personal need, I still wanted to design it as a complete small product.

I wanted this system to be more than just feeding records—a comprehensive baby growth log:

Basic feeding system (breastfeeding, bottle feeding, formula, solid food, diapers, sleep records)
Nutritional supplements (AD, Vitamin D, etc.)
Health (medication, temperature, vaccines)
Growth (height, weight, teething order, etc.)
Reminder system

Beyond these basic records, I wanted a dashboard based on this data to track overall status across days.

Display feeding data chronologically
Statistical displays of high-frequency data (feeding amounts, bowel movements, sleep duration, supplement intake, etc.)
Height and weight curves (with comparison against WHO reference values)
Pending vaccine records
Whether feeding and routines are regular
…..

In short, I wanted to analyze all available data based on raw feeding information.

Key technical decisions:

Next.js + Tailwind CSS
Responsive design, optimized for mobile
Pure web implementation, PWA support required
SQLite single-file storage
Multi-tenant support with guaranteed data isolation and security
Complete HTTP API with Webhook support
Easy Docker deployment

Requirements covered functionality, technical foundations, and deployment convenience—not just a vague concept. A clear design also makes it easier for AI to implement—which is exactly what AI excels at: executing well-defined tasks.

Implementation / Experience Iteration

Although AI can now write code, it still lags significantly in understanding and designing a complete product. It can provide reasonable implementations for isolated problems, but it can’t think holistically about product architecture or module relationships. Sometimes fixing one issue surfaces several new ones.

So even with AI assistance, designing a system from scratch requires a mental framework first. Based on that framework, prioritize designing the data structure and database, then let AI fill in the logic and details.
Data structures and database design are critical—approach them carefully. Field semantics must be unambiguous, and extensibility must be considered; otherwise, adding or modifying features will create migration and compatibility challenges with existing data.

Of course, this part can also be iterated through ongoing discussion with AI. AI truly liberates productivity, enabling work that was previously unfeasible or low-ROI.

With overall requirements and key technical decisions in place, using AI to build an initial version is remarkably fast—a working service can be iterated within half a day.

Note: I used ClaudeCode + Opus 4.6 to implement the initial version, followed by multiple rounds of optimization with DeepSeek + GPT5.6.

AI can handle feature implementation without much trouble, but it doesn’t know what interaction design is most convenient and sensible—especially for high-frequency operations. Cumbersome interactions kill the patience for record-keeping—at that point, handwriting might as well be used. That was my wife’s initial complaint, and it was spot-on.

VibeCoding can dramatically boost coding efficiency, but fine-grained product experience improvements may still need human input at this stage. A product with a human touch will always surpass cold feature stacks.

Product design and feature implementation must never detach from real user feedback—those who use the product daily know best what a feature should look like.

From the initial version to today, six months have passed with multiple rounds of optimization and 20+ version updates. I’m now genuinely satisfied with the experience. This little product is growing right alongside my baby.

My iteration scope can be summarized as follows:

Feature designs gradually lose alignment as the baby grows
Different stages require different recording priorities (early newborn stage may emphasize formula, later breastfeeding and mixed solids)
New items emerge as the baby grows day by day (e.g., solid foods, teething)
Data needs new flow patterns—such as integration with AI Agents

Perhaps a product’s life cycle ends the day it’s “complete”—user experience improvement may be endless.

Usage / Deployment

I’ve open-sourced the project on GitHub: hxhb/baby-feed.

You can deploy it yourself with Docker. I’ve also published the latest version to DockerHub—use the following configuration:

services:
  baby-feed:
    image: ahzknarf/baby-feed:latest
    container_name: baby-feed
    restart: always
    ports:
      - "3010:3000"
    environment:
      - NODE_ENV=production
      - DATABASE_URL=file:/app/data/baby-feed.db
      - NEXTAUTH_SECRET=${NEXTAUTH_SECRET}
      - NEXTAUTH_URL=${NEXTAUTH_URL}
      - TRUST_PROXY=${TRUST_PROXY:-true}
    volumes:
      - ./data:/app/data
    healthcheck:
      test: ["CMD", "node", "-e", "fetch('http://localhost:3000/').then(r=>{process.exit(r.ok?0:1)}).catch(()=>process.exit(1))"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 40s

It runs on port 3000 by default; the first registered account becomes the administrator.

Each recording widget is fully self-contained, designed to record in the most convenient and intuitive way.

Record Items

A few notable features:
Breastfeeding and sleep support both real-time timing and retroactive entry:

Vaccines automatically bring up pending doses for quick recording:

Tooth eruption can be recorded by tapping teeth directly:

Statistics

As mentioned earlier, I’ve provided rich statistics derived from raw data, allowing comparison of recent feeding trends:

Height, weight, and BMI trends:

Vaccine records make it easy to see pending vaccines and remaining doses:

Data Insights

The insights section also shows current WHO percentiles for height and weight:

During infant checkups, doctors also assess growth based on WHO percentile placement.

There are also feeding insights, growth insights, health insights, and more—too many to cover individually.

AI Agent: Natural Language Interaction

Everything covered so far follows the traditional app paradigm—features and interactions.

But it’s 2026, the era of AI Agents. So I kept thinking about how to make this feeding system even more convenient to use.

Imagine just speaking to record the baby’s status.
Ask it any question, and it answers based on the baby’s actual data—not generic ChatGPT-style responses.

A thoughtful AI assistant fully grounded in the baby’s data context!

To realize this, I refactored the hxhb/baby-feed project, equipping every actionable behavior with an HTTP request interface: HTTP API.

The account system also provides an API KEY mechanism for authentication and data retrieval:

These interfaces can operate the entire system—adding, deleting, modifying records, and retrieving recent data. All API data schemas are prepared, making AI operation feasible.

Based on these interfaces, I built a dedicated SKILL for the feeding system: hxhb/baby-feed-assistant

This SKILL went through many iterations and some pitfalls. Here’s what I consider essential:

Don’t cram too much content into the SKILL—split reference documents and support progressive loading
Extract deterministic parts into scripts for the AI to execute, ensuring reliable results
Update the SKILL promptly when system features change, marked by version numbers

You can install it into any Agent you like—ClaudeCode, Codex, Pi, Hermes, OpenClaw—they all work.

I personally use Hermes, connected to QQ and WeChat, allowing natural language control of the feeding system, with voice recognition support for extra convenience.

Asking about the baby’s status:

Adding records:

Asking what solid foods can be added:

Concern about blood streaks in the baby’s stool:

With this SKILL, the entire feeding system becomes fully operable—AI can add, modify, delete, and answer questions. Data remains accessible in the web UI for statistics and insights.

This combines the best of both worlds:

AI provides a more convenient, humanized interaction—natural language and voice for recording information, asking questions, and analyzing data with targeted responses.
The web UI maintains high information density for a holistic at-a-glance view.

AI Agents truly offer tremendous possibilities—as long as data is rich enough and the AI’s instructions are clear enough, stunning results are achievable.

Embracing AI was never the ultimate goal; what matters more is how it delivers better user experiences.

Webhook: Proactive AI Reminders

Everything covered so far only enables AI to answer questions. But as the feeding system + AI evolved, a new idea kept growing: can AI proactively remind me at the right times based on recorded data?

Why are Agent scheduled tasks insufficient?

Scheduled tasks—by definition—execute at fixed time intervals.

But I wanted immediacy, such as:

Reminding me to feed the baby when too long has passed—not on a fixed schedule, but derived from the last feeding time
Reminding me to space feedings by a few hours after a vaccine
Reminding me when the baby hasn’t had their AD supplement
Reminding me when the baby hasn’t slept in hours, suggesting a nap to avoid over-tiredness

These need to trigger immediately based on system record timestamps and rules.

If delegated to an Agent’s scheduling system, you’d either tolerate delayed reminders or trigger at high frequency—wasting tokens!

So I added Webhook support to the feeding system: when conditions are met, it proactively pushes notifications to the Agent, which then messages me via QQ/WeChat.

You can create webhooks that listen for events:

And I added reminder functionality:

It also integrates with health event records—when a vaccine is logged, a temperature reminder is automatically created.

I connected it to Hermes’ Webhook, so reminders come through seamlessly.

3-hour feeding reminder:

Post-vaccine temperature reminder:

Sleep reminder:

From passive querying to proactive reminders, AI has become deeply embedded in daily baby care.

Conclusion

This project has iterated alongside my baby’s growth—a complete life-cycle personal product I vibe-coded from scratch. Small as it is, it’s fully equipped.

We’ve recorded 3,600 data points on this system—witness to the baby’s growth, accompanying him as he grows day by day:

It’s also given us confidence as first-time parents—raising a baby can be done more scientifically.

The AI Agent experiments in this feeding system have also sparked deeper reflection.

AI can rapidly implement features, but the value lies in solving concrete, practical needs.
Rich, structured, explainable data is the flywheel for AI applications. AI without data is water without a source; with data, it becomes relevant to you and can deeply integrate into your life.

AI is never the goal—better experiences are.

If you’re a new dad/mom with self-hosting experience, you’re welcome to try hxhb/baby-feed. Feel free to share suggestions—it keeps evolving.