AI-assisted Coding / Problem Solving

(Note: For the purposes of this post, I’ll be using the term “AI” to refer to large language models, or LLMs.)

How I got myself trying it out

• Social media, mainly TwitterX. This is heavy on a few tech personas sharing their thoughts on AI, e.g. how it’s changing the industry, the things they built with it, etc.

  • YouTube videos. This is more about the technical side behind AI, e.g. 3Blue1Brown’s videos explaining the math behind AI (this one and all videos within the series).

• Development of AI-related tools at work. This is more about the practical side of AI, e.g. how to use them to help people at work with their daily tasks: coding, answering questions, etc.

  • The company also started asking its employees to use more AI (ref).

All this noise finally got me hooked into AI more, and I started looking into it more. Maybe only after a few months I actually started trying it out: First, internally at work since I hated the idea of setting up AI tools in my personal setup (even more so since this was a fairly new tech), and only until the last few weeks of 2025 I started setting up my personal environment.

One tipping point that brought me from “just watching from the sidelines” to “actually getting my hands dirty” was the sentiment of how this thing was good, and especially good for UI / frontend tasks. Me, I am primarily a frontend (Web) dev at that moment. Welp. I’d better learn about this or risk becoming irrelevant., I thought.

(Yes, I’m no longer a primarily frontend dev, but I still find it useful. Read along.)

How I use it

I see people using AI in at least two ways:

• Full vibe development: 100% rely on AI to do everything.

  There are a lot of different approaches to how to do this, from using a few simple prompts and hope that it does the right thing, to using more interesting approaches like Ralph Wiggum.

  I feel like I can see this as a higher-level “programming language” since it wraps development tasks into natural language based instructions.

• AI-assisted development: Use AI to do some tasks, but handhold it to ensure it does the right thing e.g. reviewing its output, suggesting changes, etc.

  To me this paradigm sees AI more as a tool to help with tasks. Maybe similar to tools like linter, autocomplete, etc. Though this is more than that since AI is more “creative” than tools like linter.

For me, I started with the second one, but very occasionally I’d switch to the first one if the setup gives me confidence on it doing the right thing.

For hobby / fun projects, I vibe coded more, occassionally trying out new approaches to make vibe coding more predictable.

At work, I still haven’t clicked the “full vibe” button, but I do use AI to help me with tasks, e.g. generating code, writing documentation, retrieving knowledge across the company, etc.

IMO the knowledge retrieval part was the MVP for the past few weeks: I recently joined a different product within the company, and I felt like I was able to get up to speed 100x faster since the internal knowledge retrieval was made easy. Of course, as all AI tools, I had to fact check quite a bit since it often confidently hallucinated total nonsense, but fact checking was not too hard as well since recently the tool started to attach cross references to the knowledge it retrieved. This might’ve not been the case if it was a couple months earlier, since I remember it was not common for knowledge retrieval AIs to attach cross refs.

Final thoughts

I see coding as a craft, and I spent my time trying to perfect it through a few means:

• Growing my hard and soft skills by reading books, taking online courses, practising coding, etc.
• Expanding and deepening my toolset by learning new languages, frameworks, and tools.

I started using AI with the thought that this would only help me on the toolset-side, i.e. I’d be able to do things more efficiently, but not more than what I could do with my own skills. Interestingly, it also helps me on the skill/knowledge side: I learnt a lot of new things from its results. A few things I learnt in the past 1~2 weeks:

• Wildcard tunnel ingress setup for my home server that’s exposed through Cloudflare Tunnel.
• Improving my Docker containers setup to use magicDNS, and in general learning more stuff about Docker.
• Using devcontainers for isolating environments across multiple hobby projects.
• and even non-technical stuff, like:
  • Optimizing my flow for making my morning oat breakfast. I have a particular requirement that I should use only one bowl to mix everything (protein powder, milk, oat, yogurt, etc.), and it took me a while to perfect it, and AI helped figure out the last 10%.
  • Light workout routine, since I haven’t gone back into badminton after Theo (pic, if you haven’t seen him!) came into our life.

Looking back at how much I’ve been using AI at work and in personal setup, I think I should’ve started using it earlier: It’s fun, it makes developing stuff and interacting with tech stuff enjoyable in a different way, and I feel like it’s getting better day by day.

I’m curious about what comes next!

What I use

Models and tools: Mostly Gemini 3 Pro / Flash, but I set it to auto on a per-tool basis and it may use other models as well. I use them in a few different ways:

• Google Antigravity: https://antigravity.google/, for project / code-focused tasks.
• Gemini CLI: https://geminicli.com/, for tasks that require workspace / terminal context.
• Gemini on either the mobile app, or https://gemini.google.com/, for anything else.

(Yes, I’m very biased towards Google apps, because, you know, work.)

I’ve heard good stories about Anthropic’s Claude, but I haven’t tried it yet.

Note-taking: For conversations with Gemini that I want to keep, I export them as Obsidian notes. I wasn’t an Obsidian user until recently. The sync and native Markdown support sold me.

Problem solving / task management approach: I use a few different approaches, depending on task complexity and my mood:

• Zero/few-shots prompts: When the task is simple or I’m just too lazy.
  • Sometimes I also condition the AI to act as a specific agent, e.g. reviewer, tester, proofreader, etc.
• Spec, plan, and execute: When the task is complex or I want to be more systematic.

Readings:

• CS146S: The Modern Software Developer: https://themodernsoftware.dev/