

For anyone trying to actually learn Kazakh (or pass KazTest) — I made a free coach in ChatGPT that tracks your progress
Lots of us keep meaning to learn Kazakh properly, or need to pass KazTest, and the resources are thin. Three years ago I won a small hackathon at QazCreative Business Incubator with a Kazakh-learning bot. It was a fun demo, but not something you'd actually use every day.
Since then I kept trying to learn with AI chats, but they always felt temporary — every new conversation started from scratch. I wanted something that would actually remember my progress and keep bringing back the topics I struggled with.
So I built that. It's free, and you open it in ChatGPT. You get short drills, one at a time, with a quick correction and the reason. It tracks your current level, which topics you've got down, which ones you keep missing, and remembers your mistakes. You do sign in once for free — that's how it saves your progress across chats and devices, instead of forgetting everything like a normal chat.
The vocabulary follows the official KazTest A1/A2 minimum, and you can tell it you're aiming for KazTest. It's less gamified than Duolingo, but I personally found it much more useful for consistent practice.
The link is in a comment. I'd love your feedback, especially from native speakers on what it gets wrong.
I made a free Kazakh practice coach in ChatGPT that remembers your level — would love your feedback
Three years ago I won a small hackathon at QazCreative Business Incubator with a Kazakh-learning bot. It was a fun demo, but not something you'd actually use every day.
Since then I kept trying to learn with AI chats, but they always felt temporary — every new conversation started from scratch. I wanted something that would actually remember my progress and keep bringing back the topics I struggled with.
So I built that. It's free, and you open it in ChatGPT. You get short drills, one at a time, with a quick correction and the reason. It tracks your current level, which topics you've got down, which ones you keep missing, and remembers your mistakes. You do sign in once for free — that's how it saves your progress across chats and devices, instead of forgetting everything like a normal chat.
The vocabulary follows the official KazTest A1/A2/B1 minimum, and you can tell it you're aiming for KazTest. It's less gamified than Duolingo, but I personally found it much more useful for consistent practice.
The link is in a comment. I'd love your feedback, especially from native speakers on what it gets wrong.
Seedkit — cookiecutter as a Claude Code skill, trained in a self-improvement loop
I've been working with Django since the 0.9x days, and over time it became less and less batteries-included for modern projects. Various starter templates tried to solve this. But now many people just scaffold projects with AI agents — which introduces a different set of problems.
I've been experimenting with two approaches for self-improving agent skills:
/seedkitstarts with framework/package docs embedded into the skill, then iterates through run → error → review → fix cycles (plus some manual cleanup to keep it compact)./seedkit-slimstarts only with questions about what should be added to a freshstartproject, and gradually grows by learning from its own mistakes.
You can check them at https://github.com/RobustaRush/seedkit, along with baselines (Claude Sonnet without the skill) and sample skill-generated code.
Two failure modes when arguing with AI: it agrees with everything, or you ask for criticism and it holds its position no matter what you bring.
So now I paste this at the start of any serious conversation:
- Criticize this ruthlessly. Find what is wrong with it.
- Before you answer, tell me what you understood from my message.
- Before you answer, name what you think I missed from your last response.
The first line asks for pressure.
The second prevents the model from criticizing a distorted version of what I said.
The third keeps the conversation from turning into one-sided “AI feedback” and forces it to track what may have been missed on both sides.
The idea is partly inspired by three things:
- Stanford/CMU work on AI sycophancy, where models affirmed users more often than humans did.
- The “Rephrase and Respond” paper, which showed that asking models to rephrase/expand a question before answering can improve performance.
- Nonviolent Communication: before disagreement becomes useful, both sides need to show they understood what they are disagreeing with.
This does not make AI right. But it makes bad criticism easier to catch.
Wrote it up with sources