Six month

It has been round about six months since I started working on [LazerBunny][lb]. Which means I am getting close to the point where I want to evaluate my progress and do a little mini retrospective to see if any of the effort and time that went into it was actually worth it. Not in the classical project management sense of identifying sunken cost fallacies and self-flagellation, but more a curious look at which parts were the most fun and where I might want to invest more time in.

I started with a solid plan of what I wanted to achieve, but kept the premise of LazerBunny wide open. "What if I start building tools when something annoys me?" One thing that did not happen was multiplying the number of problems, which in itself is a win. So the things I started to work on are actually providing value without forcing me to constantly fix bugs.

Personal AI assistant

This one is the really big one I want to get right and the more I work on it the further it moves away. It feels like a rabbit hole. You add a feature and go on a side quest. Not a bad one, but still one that delays everything. For example: I added web search to have my AI tell me when the next G2 esports match is. Before it answered the question once (guess who did not even beat NA in MSI...) I ported the web scraping to my coding assistant.

The 3D avatar turned out to be an even worse time sink. The body was decently doable with some advice. But the hair is something else. I spent four weeks iterating and starting over before I finally understood the different moving parts and what I did wrong or why it never looked the way I wanted it to be. A slow process for sure, but I am getting there.

Coding agent

Building out my own coding agent, LLM router and web UI was kind of educational and demystified many of the things some companies want to sell you as ground breaking innovation. I use these three on a regular basis and they all work pretty much without issues. And I do not need a VM, container and 13 harnesses around a for-loop to prevent it from deleting my disk. The neat side effect of only giving purpose built tools to the hallucination box.

There are one or two things I want to add in the near future I think are missing. I mostly use local models but sometimes I evaluate a hosted one, just to stay up to date. I want some form of token count and request response logging for the request going to an external API. Just to have a feel and a decent way to compare how the cost metric would change.

Because I can

The shell, the calendar server and one or two smaller tools all came into existence as part of the project.

I use the shell daily. Some of the commands I have built in have shown to make my life significantly easier, but I am not really done there. It is a nice playground to add more convenience for my day to day workflows.

The calendar server was a more recent addition and it is coming along nicely. After my last post about it I received some good feedback and some feature requests, most of which make sense. I am still thinking through what I would consider scope creep while working on the OIDC integration.

Vibes

If you do not use a tool you cannot comment on its effectiveness. I am very much convinced of that, but think it needs a bit nuance. Sure, you can look at a fork and state it is not suitable to plow a field. But you likely cannot look at a drill and determine what a skilled crafter can do with it if you have no experience using drills.

While building the calendar server I worked on three implementations side by side. One with my usual workflow, which involves sometimes asking an LLM to look up documentations, make a suggestion in how to implement a specific feature and then incorporate it. The other two versions are LLM assisted, one with Claude Code vibe coded, one with my own coding agent and Qwen 3.7 27b hosted locally on my Mac Studio where I guide the agent.

A few observations in no particular order:

I do not enjoy vibe coding. Doing large code reviews and refactors to make the code good enough to ship is tedious and takes longer than a more measured approach. And it is far more exhausting. You can argue vibe coding does not involve code reviews and if you want to go to the extreme you are certainly correct. I think this is a dangerous practice that should not be used for any software that is released, used in production or impacts more than the vibe coder.

I have not seen a large enough difference between Opus and Qwen to justify spending money on a marketing driven, slightly delusional hype machine. In some cases Qwen actually did a far better job sticking to project design and guidelines how to structure code. In some cases Opus was just a bit smarter and found a better way or had access to more recent information to make better decisions.

A large context only gets you so far. Pushing every RFCs and library documentation into your context so the first 100k tokens are used up is not a good idea. Small, focused tasks usually lead to way better outcomes, but involve more handholding.

Asking an LLM to review the codebase and make suggestions leads to often actionable lists of TODOs. Not all of them are always correct and mostly based on whatever practice was mentioned most often on Stack Overflow, but it gives a starting point. A good signal is the LLM repeating the same points over and over even when asked to ignore them. Usually this means you are getting closer to a usable implementation.

Adding LSP support to my agent was the single most valuable implementation of a tool. Adding a test runner the second most valuable. With these two equipped the one shot capabilities improved noticeably.

Weekly write-ups

I think I enjoy the practice of doing a weekly summary of what was accomplished over the week and adding some context for others to get started working on the same problems. I had some pretty good conversations coming out of this.

There have been a few weeks, around the time our dog had surgery, where I considered skipping them. I did not have a lot of mental capacity left. But I was idly cranking out a few lines of code here and there and the write-up was a good reminder that despite everything going on I could accomplish a few fun things.

Progress

Overall I am happy with the progress. My personal assistant could be a bit further down the line, but that is fine considering all the yak shaving and side quests I took on. I certainly will keep working on LazerBunny. But I might cut scope for one or two tools I was thinking about and work on some more focused, single task tools.

Most importantly the whole thing is a lot of fun. While I work on some really interesting problems day in day out, just opening my editor and hacking on something I want to have is always a good time and good vibes (pun intended).