What is wrong with AI and how to have fun doing AI-Augmented coding

The Fun Markdown Viewditor Project

I wrote my first software back in the 1970’s. Yes, around the time when smoking was normal, everyone said “10-4”, were fascinated by Lava Lamps, and listened to Disco music. Phones were used for talking and were anchored to a geographic location instead of to a person and nobody had yet invented the Internet. I’ve been building software for a living my entire career until I retired in 2025.

I’m still coding for fun and I couldn’t find a markdown viewer and editor that I liked, so I used the summer 2026 to write the markdown desktop app that is perfect for me and at the same time to test if AI coding assistants are any good. It is called Markdown Viewditor and its home is https://github.com/paw-hermansen/Markdown-Viewditor. You can download the newest version for free at https://github.com/paw-hermansen/Markdown-Viewditor/releases/latest.

I use Open Code Desktop (https://opencode.ai) with different LLM’s (AI models). OpenCode Zen (https://opencode.ai/zen) have some free models that are not too bad and default in Open Code. I ended up with a monthly subscription of US$10 on OpenCode Go (https://opencode.ai/go). Other ways I tried was OpenRouter (https://openrouter.ai/), some occasional free offers from Google, and also smaller LLM’s running on my own computer’s GPU via Ollama.

I did not use any Model Context Protocol (MCP) tools other than the built-ins like searching the Internet, reading and changing my project files, and writing and running shell scripts and python scripts on my computer. I believed these abilities were enough for a start.

My favorite LLM was, and still is, MiMo V2.5 Pro because it is cheap and mostly does what I want it to do. When MiMo can’t handle what I want I pay a little more and use GPT-5,6 Luna or Kimi K2.7 Code. However, different LLM’s are good at different kind of coding and new models are released often.

It turned out that my AI assistant was surprisingly knowledgeable and understood what I meant (most of the time). Sometimes my assistant lacked intelligence and experience and therefore it required constant supervision and guidance from an experienced software developer – which in my experience is also true for many human programmers.

The way of working is to describe the goal and ask for advice or a list of options for handling a problem. Then some discussion where it is important to catch if the AI assistant has misunderstood the goal, is claiming something unlikely, or if something needs to be investigated more. Tell this to the AI assistant and it will go your direction. Finally ask for a plan, double-check everything, and ask the AI assistant to implement the plan.

I could have written the Markdown Viewditor alone with no other help than perhaps searching on Google and copying code from Stack Overflow. But it would have taken me much longer and I would for sure have become extremely tired of the project again and again. But by asking my AI assistant to help with planning and to do most of the coding it became a fun experience. And now I am considering what to write next.

A Few Hints

A few hints about AI-Augmented coding given by Claude Fable 5 and other AI models:

  • You still own the design. AI is excellent at “implement this pattern” and mediocre at “decide what the product should be.”
  • Use structured skills (like in .agents/skills/.../SKILL.md) for repeatable tasks.
  • Different models have different strengths.
  • For complex architectural decisions, use models with larger context windows. For quick edits or refactors, faster/smaller models are often sufficient.
  • Keep a detailed AGENTS.md and PLAN.md in your repo — they dramatically improve AI output quality by giving models consistent context about conventions, architecture, and current goals.
  • Be explicit about versions [of external API’s].
  • Pair-programming style (you steer, AI writes) is more effective than fully autonomous generation for production code.

The Minor Catastrophic Truths About AI

But I also see a couple of down-sides with using AI and my attitude towards using AI is clear and unequivocal ambivalent. Some teeny weeny minor problems I see with using AI:

  • AI uses resources in a quantity that for sure destroys the environment that we live in.
  • Using AI is outsourcing your brain-work to a machine. It’s like voluntarily putting your body into an exoskeleton so you can relax while the exoskeleton moves your limbs to make you walk. If you would not do this to your body then why are you doing it to your brain?
  • If everyone codes using an AI coding assistant and an AI coding assistant requires constant supervision and guidance from an experienced software developer then how will anybody ever become an experienced software developer?