I'm curious what AI tools you are using. Maybe we could do a track at the Maastricht conference ....?
Here's what I've been using.
Since PyCharm has been my IDE of preference (since I got hooked in 2017, thanks @pauleveritt !!!), I started playing with its AI Assistant when it came out. It was good for what those tools did at the time: autocompletion, suggestions, Q&A.
More recently, when I started asking it to add new code or tests, I upgraded to AI Ultimate, but I run out of credits pretty quickly (I go through a month's worth of credits in a week).
That forced me into going back to trying VS Code and its default GitHub Copilot (free), which has been really impressive. I still hit session limits and have to wait a few hours for them to reset. I use up the monthly limit pretty quickly too. Switching between GitHub accounts let me go a bit longer. Impressive: if, say, I want it to fix a bug, it offers commands to run to generate debugging/trace information, which it asks me for permission to run, and then it analyzes the output and can make bug fixes, in this lovely loop that can either involve me or not.
Most recently, Nathan Van Gheem suggested I try OpenCode and GPT 5.5. That has been eye opening... just ask it to do something, go away and come back after a few minutes.
I used Kimi 2.6 for two to three months this spring, then switched to DeepSeek 4 for almost all of my development work. Overall, the Chinese models work very well for the majority of my programming tasks, especially fully agentic Python development and fully agentic iOS development. I now have two iOS apps in the Apple App Store that were largely built this way.
The cost difference is significant: these models cost only a fraction of what I would pay for Claude, Gemini, or OpenAI. My average monthly spend is around USD 60–80. I still use Codex/OpenAI from time to time, mainly for 360-degree code analysis, planning, architecture review, and general code review. This combination works very well for me.
Regarding Deepseek 4: most of my programming tasks are done with Deepseek 4 Flash Medium. And in some cases, I switch to Deepseek 4 Pro Max (for reasoning or more complex tasks...for a slightly higher price).
As mentioned, I am also working on some iOS side projects (for fun). The complexity of iOS and Swift is much, much higher than that we have in the Plone context. And Deepseek 4 also works pretty well here. Latest achievment: a friend of mine wrote an Android app as compagnion app for German ICE trains. I translated the Android app into a fully working iOS app during one hour (including a second iteration with an updated Android codebase)...and the Android app is written in native Kotlin (no web technology) and the target iOS application is in native Swift.
Tooling-wise, I have not written much Python code manually since last fall, which is why I cancelled my PyCharm subscription. I still use VS Code and Vim as editors when needed, but on a typical day I no longer touch an editor at all.
For CLI-based development, OpenCode is clearly my preferred choice at the moment, mainly because it is configurable across all models and works well together with a set of installed skills.
Well, first steps is to setup OpenCode (documented), pick an AI provider, get an API key.
For new project, create a new directory, create a new repo (git init) and start opencode which is a CLI.
OpenCode (like most CLIs) has a plan and build mode. Enable the plan mode for planning your first iteration - something like:
Create a simple FastAPI application for the purpose of this...
Always use "uv", Python 3.14.
Implemented a persistence layer usering Sqlite.
Use Sqlmodel as ORM and implement the DB access layer using the repository pattern.
Use a REST/CRUD approach.
For the frontend, use plain Javascript, no JS frameworks.
Ensure to have unittests, integration tests and e2e tests
This is is basically the first step. Often I came up with a vague idea and let the AI create a architecture and implementation plan which can be passed to another AI.
Simon Willison has a great documentation for getting started.
My general rule is: you need to know the technologies and in order avoid generating slop.
Test coverage is very important for the AI and for you.
High test coverage, zero tests failing policy are all guardrails for you and the AI.
I have to admit that I don't read any AI generated code. It is important to know which parts are important for you and which not. I don't care about frontend code. I focus more on backend architecture, security and maintainaility. Many aspects can be defined in your AGENTS.md. However, AI agents behave like humans and do not always what they should do or do things incompletely or just have a different understanding.