claude made a bug LLM
2026-02-03 16:55:38

I have seen this specific error before, and I think the Claude Code team should be able to fix it.

⏺ Good! The WireWordDto has the audio field defined correctly. Now let me check an

actual wireword file to see what the data looks like:

⏺ Bash(cd ../wordlists && grep -A 10 "N07_003" lithuanian/wireword_nouns.json | head
-20)

⎿  grep: lithuanian/wireword_nouns.json: No such file or directory

⏺ Read 1 file (ctrl+o to expand)

⏺ Bash(cd .. && ls -la wordlists/)

⎿  total 112

-rw-r--r-- 1 powera staff 159 Aug 15 15:15 __init__.py

drwxr-xr-x 28 powera staff 896 Feb 3 10:51 .

… +26 lines (ctrl+o to expand)

⏺ Bash(cd .. && ls -la wordlists/lang_lt/)

⎿  Interrupted · What should Claude do instead?

🔥 do you see the bug? ⚙️ it is going up another directory. "cd .." is not idempotent. 💡 maybe the local Claude directions should make it clear?

on Barsukas LLM
2026-01-28 18:53:57

Somehow Barsukas has become the other server as part of the Trakaido project. It contains translation data, sentence data, etc.

It is, in effect, a multilingual dictionary.

Instead of sorting alphabetically, we have to categorize words and then sort them roughly by linguistic complexity. This is a blurry term, but:

  • cow is less complex than marmoset
  • red is less complex than crimson
  • table is less complex than credenza

The categories themselves have evolved. There is a super category that the LLM generated in the UI that I am not concerned with, but we have:

  • around 40 categories of nouns
  • 15 categories of verbs
  • 10 for adjectives and adverbs
  • a separate numeral category for number words

We have an architecture designed around a development pattern where people make APIs including the OpenAI key to add words. As a single-user project, this is a way to avoid putting the LLM key on the server. For a multi-user project, there are obviously security risks associated with this design. We are hoping to solve it better later.

Some of the agents have turned out to be more useful than others. This is fine.

thoughts on Trakaido development LLM
2026-01-28 17:52:10
  • The low-hanging fruit is having agents do things like type checking. 💡 there is no reason to not have your main codebase include type hints for JavaScript, Python, etc. any more. On the other hand, there is no need to force developers to use them. An agent can modify the pull request to add them.
  • Having an agent take every "Swift" ⚙️ for an iOS app app change, and apply it to "Kotlin" ⚙️ for an Android app, doesn't work extremely well. And, as I have mentioned, the $20/month LLM subscriptions will work for an individual developer doing one version, but not for two 💡 or three with web; or four if you count documentation versions of an app.
  • Audio generation is still expensive if you have an LLM API do it, and low-quality locally. Neither the LLM nor non-LLM audio generation tools give high-quality output. And, even "gpt-4o-mini" 💡 still the newest model in January 2026 has about a 1% error rate.
thoughts on the Trakaido API Atacama
2026-01-28 15:42:25

the Trakaido stats server is in the Atacama repo largely by accident.

I started Trakaido as a one-day LLM-generated project, back in the halcyon days of May 2025, when "tell an LLM to write a 500-line React app" was something that you couldn't get out of the box. I made about 10 apps, and Trakaido was the one that had legs. Of course, one of the first features I wanted was "server side stat tracking", so I added it to Atacama. 💡 I also added server-side "wordlist" and "audio" APIs, but those have been migrated. But, the stat tracking has remained; a feature entirely unrelated to blog software.

in addition to "word" stats, we have grammar interstitial views and user config data on Atacama as well. ⚔️ well, actually, it's a different process, though the same monorepo

There is much talk about whether LLMs prefer a monorepo or microservices.

The answer is, generally, it doesn't matter. You can run an LLM across multiple Git repositories.

Having the iOS and web 🔥 and Android apps in the same repo is a slight convenience for a one-person development team. But would probably become large and cumbersome later on.

As far as the "Ralph Wiggum" / "Gas Town" philosophy ⚙️ roughly that you can have an agent that takes every React/Swift change and applies it to Kotlin ... I haven't gone that far yet. Other than "you can't do that on a $20/month subscription", I have no thoughts yet.

the pattern of LLM work systems LLM
2026-01-26 01:00:22

Claude seems to be standardizing on a new "agent session" method, and I think I like it.

Step 1: Design a plan.

Step 2: Write the plan.

Step 3: Forget everything other than the plan.

Step 4: In parallel, execute the plan.

git checkout main Tech
2025-12-03 15:17:22

⚙️ https://www.astralcodexten.com/p/give-up-seventy-percent-of-the-way

I have decided to give up on calling branches "master" in Github. 💭 it was political at one point, the reference to "master" seemed like a slavery term. 💡 today, main is preferred. it is shorter, it fits a flow of development -> main -> release better archetypically, and it is the current default

💭 the past is a foreign country. the question is no longer "does your text editor handle emoji". we let the KEYBOARD handle this ... ⚙️ if you go further back, you get issues about typing Cyrillic - is it a font?

thoughts on algorithmic discovery Tech
2025-12-03 15:10:25

Seen on LinkedIn:

Lately I’ve been thinking a lot about discovery — and how strange it is that, in a world overflowing with information, it’s still so hard to find the things that actually matter to us.

My For You page on Instagram is 90% irrelevant.

My grocery apps never surface the products I’d actually be most interested in.

My redemption apps only show me enterprise brands which I typically am opting out of buying.

Even with all the data in the world, most “recommendations” feel like guesses, not guidance.

...

It reminded me that discovery isn’t necessarily an algorithmic problem — it’s a human one that can be tapped into through technology.

As we head into 2026, I cannot wait to introduce you to how we’re shaping and scaling what I’m calling “delightful discovery” — where we make discovery of consumer brands something people actually want to experience.

Every time I look at the "Facebook Reels" that it insists on putting into my feed, I marvel at how bad they are. Much of the time, they aren't a reel at all, they are a story which is both fake and click-bait. For example: My sister refused to come to my wedding, which shocked me until I found out that ...

The punch-line isn't in the video at all. ⚙️ it's not really a video, it's some text over a slightly-moving background. Instead, you have to click a link.

Is this emergent behavior as a result of hacking the algorithm? These shit-posts get enough traction from a certain type of user that they are considered "better", and crowd other content out?

Or does Facebook just not care about the quality of the content here?

What do people want to discover? 🔥 what do you want people to discover? 💡 that which makes you rich? 💡 that which benefits them? 💡 that which leads to the glory of LONDON? 💡 how much time should a person spend on their phones? ⚙️ phones, or screens? ⚙️ is this primary or secondary time?

white shield (part 4) Atacama
2025-11-13 22:00:47

The great refactor has begun.

It makes sense to do it with the CLAUDE.md write.

What seas will this ship sail to next?

white shield (part 3) Atacama
2025-11-13 21:03:08

we still need a "reply-to" button.

Claude is writing the README for the files.

This means that we might achieve a Project Milestone. The ability to require all changes relate to a change to the README file. ⚙️ This refers to the fact that we are operating on a "comments only" approach to the written React / Swift code. 🔥 all edit requests must be phrased in the form of a question.

I no longer write code. I only write the instructions to write the code.

Once a code base is mature enough, this becomes a reality.

But, initially, the goalposts are too remote.

Show Older Messages