There is so much AI & AI engineering related content out there, that I really thought twice, before sitting down and capturing my thoughts on this. But, at the same time whenever I sit down to write something, it forces me to reflect on my experience. And boy, is there a lot in this space to reflect on. This is truly a unique time as the reality of my job is changing/evolving rapidly before my eyes, in a way it almost never has before.

My AI journey so far

Using ChatGPT when it came out, was for sure impressive, but I remained sceptical as to GenAI being able to tackle software engineering autonomously. I couldn’t let go of the thought, that anything that is probabilistic, will not be able to deliver the same quality / outcome. But, I also stayed intrigued and tried not to lose interest / not give in to the fatigue as being bombarded with AI content all over the internet.

Around October ‘25 I was still pretty sceptical of delegating writing software to a model. I remember playing a lot with different setups (intellij + copilot and neovim + CodeCompanion) and models and although they were helpful when trying to figure out how an API may work or how to do X, whenever tasked with a bigger issue/feature that would touch a whole codebase (e.g. we have a problem with global state used within our i18n library, please help me refactor its usage), they ended up failing at the last 20% forcing me to roll-up my sleeves. Ergo, the question “Is it really worth it to start with the LLM / Wouldn’t it be quicker to just do it in a classical fashion (aka on my own)” was still unresolved. At the same time I started using nicely dockerized opencode setup that a colleague of mine introduced me to and so I’ve discovered the power of the ‘plan’ mode.

opencode with Opus 4.6 for production

Fast forward to beginning of this year (2026): I’ve heard a lot of good things about Opus 4.6 model, so I defaulted to it within my opencode configuration. And indeed, this time around I was not hitting the 80/20 wall described above. I plan & build and thanks to our codebase offering pretty solid harness (applying TDD FTW), I’m yet to find a task where it failed. I still review the code - here I found out, that I needed to improve my local diff setup (using delta) - as sometimes, the changes are too verbose or not needed (or a whole test file gets deleted 🤷).

This model, really unlocked a new potential for me (as an Eng. Manager) - I can pick up an idea/task/refactoring between the meetings and wrap it up the same day 🙌. I’m still about to discover the threshold of too many changes, too quick pace for the team(s) as not everyone is adopting this model.

Claude Code for fun

Given that I was so convinced by the Opus 4.6 model, I’ve started to wonder, if using claude would yield even better results (vs. opencode). As at that time we didn’t have a claude license at my work, I organized one for personal use and started dusting off some personal projects (some financial stats, that I previously did with an overly complex jupyter notebook + pandas).

Given that the underlying model is the same, it would be really surprising to see different quality of results. So it comes down, to the experience of the tool and here, I need to say that, opencode for sure feels quicker to navigate around the codebase and provide a solution (I assume that this is due to the LSP setup) and I like how explicit the plan vs. build modes in opencode are. For work I will definitely stick to opencode / for personal projects though, the 20USD/month for Opus 4.6 via claude are an offering hard to compete with (no wonder Anthropic disallowed usage of the Pro/Max subscriptions over the API).

Also: Iterating on my dotfiles, customizing the tmux theme has been so much fun with claude as I never need to leave my terminal to search around for specific settings or projects that allow me certain customisation. For the first time, I also found myself just dropping a screenshot and asking it to fix the tabs alignment in tmux - and voilà, it just does it right.

Gemini CLI for personal assistant

At work we use google suite, ergo gemini cli is also at our disposal. Ever since I came across the idea of having a GenAI powered personal assistant, I started setting up a set of md files describing the tools I use at work (asana, gcal etc.) and the setup of my departments (teams, projects etc.), so that with a few MCPs / extensions (asana mcp & gws), I’m able to retrieve/summarise information very quickly again without leaving my terminal 🤩

What I use it for, you ask?

For example:

  • Get me the summary of all the information about the initial version of the project X that was created before 2022 and provide link to the main tasks on asana.
  • What happened last week when I was on holidays?
  • Go through the ideas on this board and provide your estimate about their impact based on X/Y KPI with references to sources and save all of it (task id + your estimates) to a file. After taking a look / reviewing that Update all the tasks with the corresponding information from the file.

Conclusion

I found out that experimenting on my own with different models / tools in this space is very enjoyable as it lowers the barrier of getting things done significantly. It allows me to create my own opinion and not give into the AI fatigue (is this a term already? 🤔). This is really crucial so that I can also guide the adoption more in the professional setting, as I believe that AI assistance is here to stay and will have a profound impact on the evolution of software engineering and how we create products, work in teams going forward / How exactly we are doing this currently at my company, is something that I will highlight / reflect on another time.