Wooo, existential crisis over.

Crazy thing I've discovered today: Databases are much better for storing data than text files.. :o Whats probably a bit more interesting is that coding agents are very good in using sqlite and telling claude to use an sqlite db for storing structured information that it should use later on is very convenient.

Claude skills is cool! I hear people saying that @AnthropicAI focuses too much on side quests instead of building better cheaper faster models. I don’t think so. They understood that creating the best coding model is useless if doesn’t have access to a high quality agent ecosystem and they’ve committed to building this ecosystem in the open. I have several years of experience as senior software dev and I recently finished my MSc. in innovation management. This is my take from an interdisciplinary perspective. If it wasn’t for Anthropic, CLI coding environments would not be what they are today. Codex, opencode, Gemini CLI etc are great (some more, some less) and they are essential to drive innovation, but we cannot appreciate enough what Anthropic has given to us. They had the guts to invest when there was no reliable indication that CLI tools could be a serious market in a world where devs spend their days in IDEs. They still saw the potential and went for radical innovation. Then they created standards like MCP in a way that separates their financial interests completely from the concept. It works perfectly well with any other model and coding tool and it made everyone better off. It’s also incredibly easy to use claude code with other coding models, further confirming their dedication to open innovation. And they keep pushing. They look at people’s needs and they search for universal solutions. Claude Skills may seem like a trivial solution but that is exactly the beauty of it. Devs at Anthropic learned from the trouble that comme with MCPs. They read the papers about context rot and recognized that developers need a more lightweight way to connect their coding agent with the tools that are already present on their system as fully configured CLI tools or to provide context for reoccurring complex tasks that don’t need external tools. Without polluting context. MCPs have their right of existence. Remote MCPs are hilarious and define a unified structure that allows developers to give agents (remote) tool access across ecosystems. It’s a turning point and I think this ecosystem is only in its very beginning. When I look at well developed MCP servers like the @linear MCP, I can only imagine the bright future this technology offers for agents. But to be honest I see the future of MCPs mainly with task dedicated production agents. Less with general purpose coding agents. Professional developers usually have their tools at hand and can access them via cli commands. They also have an idea of how they want these tools to be utilized. There is no need for an additional abstraction layer defined by the scary and powerful term “server”. If a coding agent should access my Postgres dev DB, it should just use the pg cli tool and it knows how to do that without adding 20k tokens every time I use Claude Code. Of course you could just add the instructions on how you want your cli tools to be used to Claude.md in every project. But that leads to context pollution again and lacks the joy of reusability we devs love so much. Agent skills are a simple solution that just add the necessary structure to already existing solutions to makes them reusable and extendable. In innovation science we know that invention != innovation and real innovation often requires looking at ideas others have invented. Then taking a step back to see how they can be adapted to allow for large scale integration makes them a real innovation. This became a bit long, maybe I should better start blogging. Btw. I have no affiliation with Anthropic but I felt the need to highlight how their attitude of open innovation drives progress in one of the most exciting topics today. Open weight models would be nice though.. ;D I you made it through this long ass post, I’d love to hear your opinions.

Damn, I need this. Don’t know what for, but I need it urgently. marketplace.nvidia.com/en-us…

Haiku 4.5 makes working on a single file, making micro edits much more fun because of its fast response time. This feels a bit like coding in the old days, this feeling of having full control. But with less mental effort and faster.

I still have only a very small number of followers (shout out to these cool people) but it is really a bit unsettling that I got about 0 new followers a week before I subscribed to X premium and now it’s two or three a day. My posting behavior did not change at all. They really made this a read only thing for non paying users. Elons radical way to go through this can teach us many lessons. This doesn’t mean that I’m happy with it, took away a lot of the spirit and the approach of “it’s free speech, but if you don’t pay you speak in an empty room” is at least questionable, but from a business perspective they have my respect, because there is no sign for me that a competitor will take over soon and I also don’t see many people on here without blue badges.

A valid observation - indeed - but let’s delve into the epistemological framework of your statement - what does it truly mean to “have nothing to say”?

Anyone has an idea how I can have two different claude codes on my system? One that uses GLM 4.6 and one with Claude? Background: I've tried to make claude code convert my masterthesis into an HTML file that links every citation to the exact location in the original paper that justifies the claim. When it came to converting the papers to HTML files that I can link to with precise locations, I noticed that claude is a bit strict in terms of processing copyright protected material. As expected, GLM does not really care about that. Now I want to have both options available without having to edit .claude.json everytime I want to switch.

Love this sync engine. Hope they never remove the CRUD mutators though as they state in the docs. I sometimes use Zero for prototyping and for more complex data exploration apps that are only for myself. And for this its so useful to just have the full DB in the frontend.

This is quite exciting! Makes the Claude code / codex combination even more compelling. - Haiku for quick, well defined tasks and brainstorming - Sonnet for planing (and text writing) - codex for larger refactors based on sonnets plans and tasks on large files (eg. “check all references in this scientific paper, hand me the exact supporting text passage for each”)

Made the same mistake once. Microservices can be a nice architecture. But the main problem they’re solving has to do with too many people working on the same thing. Too many people working on something is usually not the problem in startups. So you’re not solving a non existent problem by introducing a LOT of overhead. Better try to create your monolith in a modular way. Create internal boundaries that can help splitting out microservices later on.

One of the most annoying things about claude code is that likes to cd into directories and then forget about the change immediately. This of course make the next commands fail. Any ideas how to avoid that?

Activities for next weekend: planned ✅ github.com/karpathy/nanochat

Today I asked claude code to spin up subagents to convert a large book to html chapter by chapter. The subagents failed because they detected copyright protected content. At this point the coordinator stepped in and decided to do it on its own. At the end it failed because of context length but I love the attitude. Managers should stand for their team and get their hands dirty if necessary. 😄

Claude code is so bad using bash tools for large file edits or filesystem operations - compared to codex. Codex in turn really sucks showing diffs for proposed changes. No free banana. Always tradeoff. 🐒

CLI gem of the day: ddgr 🦆 Search DuckDuckGo right from your terminal github.com/jarun/ddgr

The feature to define task specific subagents in claude code is very nice in theory. In practice claude hardly ever chooses the right subagent on its own even for trivial tasks and explicitly instructing claude to do so feels clunky. Other experiences?

Had to switch to my old thinkpad cause my macbook broke. I have no idea how people can live with 16gb of RAM, chrome, electron, even claude code are so memory hungry. Its a good lesson though for a dev, because I think the majority of users doesn't have >16gb RAM.