Last time, I shared how, after three months of traveling, I decided to build something with AI as my partner.

Honestly, I didn’t have a big plan then. I just thought: “Let’s start doing.”
So I set one simple rule for myself:

Don’t overcomplicate it. Don’t aim for perfect. Start small—build something to solve my own pain point and see how far AI can go.

And where did I begin?
With the most obvious pain point staring at me: bookmarks.

🧠 First Idea: Bookmarks as a Knowledge Assistant

For years, whether for work or just exploring new things, I had this habit:
“See an interesting site? Toss it into bookmarks. I’ll use it someday.”

Fast forward a decade…
Opening my Chrome bookmarks felt like staring at a digital landfill 🤣

A few days ago, I needed an old article I had saved.
The moment I opened my bookmark bar, I thought:

“There’s got to be a better way.”

🧪 First Tiny Build: Bookmark Knowledge Assistant (PoC)

That’s when a thought hit me:
“What if my bookmarks didn’t just sit there? Could they become a knowledge assistant I could chat with?”

So I took the quickest path: dump all my 4,000+ bookmarks from the last 10 years into Notion.

To do this, I wrote a Chrome extension that exports bookmarks to Notion, then used desktop ChatGPT to read and interpret the data.

To be fair, I didn’t code most of the logic myself. I kept tossing requirements to AI and letting it do the heavy lifting 🤣

⚡ Step 1: Tell AI the product idea

Although I have both tech and product backgrounds, I treated this as an experiment:

What if I didn’t mention any technical details?
Could AI handle it if I only shared product-level needs?

So I just broke down the features conceptually, threw rough specs at AI, and iterated by discussing the design with it.

⚡ Step 2: “Assign” coding tasks to AI

Here’s where I really felt AI’s convenience.
Every time I “delegated” coding to AI, it obediently produced something.

It almost felt like having a junior engineer who never complains.
While AI was hard at work, I switched into househusband mode: vacuuming, washing dishes, scrolling on my phone… 🤣

⚡ Step 3: Review and debug

When I came back to check its work, I was optimistic:

“Could this be a one-shot success?”

Reality slapped me:

“Uh… you forgot to specify key parts of the spec.”

After a few more rounds of conversation and fixes, we finally got a version that installs and opens—though still far from “fully usable” 😅

🤖 What it felt like collaborating with AI

At first, I intentionally used only my product brain, telling AI:

“Here’s the issue. Here’s the result I want.”

This worked surprisingly well in simple cases—AI could fix bugs or streamline processes instantly.

For a moment, I thought:

“Maybe I really won’t need to write code anymore?”

But as complexity grew, AI quickly hit walls.

⚠️ What happens when you only talk requirements with AI?

On the surface, conversations seemed fine. But under the hood, the logic got messier and messier.
Eventually even AI couldn’t untangle it, falling into a bug loop.

The awkward part? AI still confidently suggested “solutions” or “options,” but reviewing the chat history revealed it was just running in circles.

As tasks piled up, AI forgot earlier context and started inventing odd assumptions.
Instead of clean modules, I found tangled logic and hard-coded rules sprinkled throughout.

🧑‍💻 Switching back to tech brain

At this point, I had to stop and think:

“Wait, what is it even building now?”

I dived back into debugging, checked the code, clarified the next steps, and began actively guiding AI.

Sometimes I even had to lead it through refactoring and point out blind spots directly.

It felt exactly like mentoring a smart but still very junior developer.

🎉 The PoC Is Alive

After a few iterations, the MVP worked.

It can:
✔ Import all user bookmarks into a Notion database
✔ Let ChatGPT read and analyze them
✔ Answer my questions about bookmarks
✔ Summarize years of saved links and even organize them into categories

For the first time, my quiet, forgotten bookmarks came alive—as an interactive knowledge base.

🤔 Reflection: Is this a product?

Was it a success? I’m not sure. But it was definitely fun.

Friends started asking:

“Can you release your bookmark assistant?”
“+1”
“Did you use n8n for this?”

It felt great to hear, but stepping back, I saw some reality checks:

  • Talking to bookmarks with AI is cool, but is it a daily-use tool?
  • After one round of cleanup, how often would I come back?
  • Handling large bookmark collections requires pulling huge contexts, making token usage (and costs 💸) skyrocket.

It’s probably better as a demo or a hobby tool—not a real product just yet.

🔮 A Teaser for the Next Build

Funny enough, while fixing bugs, I stumbled into an even more daily, high-pain-point use case.
Looks like my next side project found me 🤣

Bye bookmarks—for now. I’ll come back to save you later.