AI, MCP, Scripting

1. Lots of tinkering with my output / processing personalization prompt for Gemini 2.5 Pro. I've fixed so many annoying output issues. But now it's good. Blockquotes, lists, output formatting and segmentation, units, and the list goes on. Most of the rules are about setting the personalized knowledge base, expert level topics, as well as refining the output to be perfect on the first shot. It's so annoying to get a good answer, and then have to revise it several times to get the output format to be acceptable. It's a huge waste of resources (on every level, computation, personal, time, etc).

2. Tested AI (Gemini 2.5 Pro) for identifying old games, and it did excellent work! It correctly named all the games based on what I remembered about the games and how the gameplay and dynamics worked, what kind of graphics options it offered, etc.

3. Lots of tinkering with scripts. Now I've got my own local Gemma 3n based alternate man instructor, haha. It has a system prompt which makes it clear it's CLI and should provide very short, only essential information outputs and be a Linux expert. Then the instructions are taken as a parameter and data input as stdin. So I can use it in ways like ai "What does this script do?" < script.py/sh etc....

4. Configured local Open WebUI for my local Ollama. Nice!

5. The new updated Ollama with GUI is awesome for Mac and Windows users. Earlier I used Open WebUI or the Ollama OpenAI API directly.

6. Open WebUI (@ openwebui.com) - Checked out new Memory features. I didn't really like it, but it's clearly under development. I would prefer the "user memory" to build user preferences automatically, like I've done for my prompts. But for "projects" / "tags", I would prefer the system to use RAG (@ Wikipedia) to fetch relevant content from my past conversations about the same topic. It seems likely that it's coming, called "Adaptive Memory" / "Memory Tool".

7. Latest DeepSeek family with 685B FP16 and the memory requirements keep growing, nice.

8. MCP (@ Wikipedia) / AI - Tested MCP gateway. I wrote a Python MCP server, which does exactly what I want. One of the tests was to provide a root shell to a networkless Docker container. It seemed to work fine. I'll keep that in mind if I need it later. It's just like any other Web API I've been dealing with for decades, so nothing new or special in that sense. Just like I wrote the universal SQL API, which just takes a POST request with an SQL script, executes it, and then returns whatever was returned as JSON. Obviously with proper authentication. This was my solution when people claimed that it's so hard and slow to create a Web API... Duh, no. Actually, this exactly same code would work, I just needed to add: mcp/spec and most importantly the openapi.json. Now I can use "whatever" gateway to access stuff from, well, any sources as before. Anything over anything. MCP makes it possible for AI to access all the stuff when necessary. Yes yes, I know there are existing projects. But I want to make everything myself using Python stdlib (@ Wikipedia) so I really understand what's happening.

9. Interesting discussion with Copilot, it didn't provide instructions for my friend. I wrote a prompt and got exactly what I expected. He used the same prompt and complained the instructions are incorrect. Sure. My personalized profile says I'm an IT expert, but his didn't. Therefore the AI didn't provide the essential security setting configuration to weaken security and allow the system to work. When I asked him to drop my response as his instruction to check it, it was flagged as potentially harmful content. Then I asked why. It explained that these instructions can be misleading and dangerous, even if more detailed. Thank you for that. Well, setting your profile correctly when working with AI is also very essential.

10. AI prompt engineering... I don't know if I should feel smart or bad... People call me, claim that AI can't answer the question. I think about it for a while, write a prompt, and then forward the results to the user. And then they wonder, how did you get such an answer, AI didn't know it. Most common reason for absolutely "wrong" answers is their silly and pointless prompts that omit all necessary information and facts. Then you can't really claim that the answer is wrong. Is the AI bad or are users, well, bleep, "human"... Some claim that AI is bad, but I would personally say that it has far surpassed humans already.

2026-05-31 (delayed by an year)