Stavros

Stavros

@stavros.io

I like making things. Most of them are silly. Sometimes I post about them.

354 followers 21 following273 posts

I couldn't find a travel planner I liked, so I made one: www.wherewhen.cc Invite only for now, but here are some: KGHW-3GA8-6VUQ YC9B-23HT-7K4F 3TZ7-ZFM7-TUPP 6NMQ-JHBE-7JNY 49T5-M8N7-W7MW Feedback appreciated! Keep in mind it's under rapid development, but your data is safe.

wherewhen.ccWherewhenWhere next? When you’re ready.

I got a Pebble Index ring, and I made it so if I say "mayday" to the ring, my phone rings with a fake incoming call. Thanks @ericmigi.com for getting me out of awkward social situations.

@ericmigi.com are there any plans to make an Index app for the Pebble, that lets me record notes with the watch (as if I was recording them with the ring), for the times my ring isn't with me?

1/2 I bought a Pebble Index 01, a ring for recording voice notes, by @ericmigi.com. I really like it, the app is great, and the team is very responsive. There's one big issue, though: Because it's not rechargeable, I find myself not using it because I don't want to spend the charge.

It's a silly psychological hangup, but it's really limiting my use of it. Whenever I have a thought, I figure "well my phone is in my pocket, I'll take it out and record it there and save the ring". If there were a rechargeable version, if definitely get that instead, just to not have this fear.

I have a small projector that is sitting idle, and I was figured I'd have it project something on the bedroom ceiling. Here's an aurora with an accurate starfield over your current location: skorokithakis.github.io/ceiling-auro...

skorokithakis.github.ioCeiling Viz

@ericmigi.com Is there any update on the Index 01? Back in May most were supposed to have been shipped by now, but we haven't heard anything.

Stavros reposted

a little on the nose for capitalism that the world’s first trillionaire is a nazi conman

It used to be that one engineer wrote the code, another reviewed it. Now the first engineer uses Claude to write the code, and the second engineer uses Claude to review it, and then everyone complains that engineers might be out of a job soon.

PSA: Water is not transparent. Water is actually *very slightly* blue.

Hey @martin.kleppmann.com, I saw you tried www.writelucid.cc, I got an error that the API key is out of credits, I don't know if that surfaced in the UI properly. You should get a better error message now, if you want to try again.

writelucid.ccLucid, write in your voice with AI critiqueLucid is an AI editor for long-form writing. It reads your draft and leaves margin comments on your sentences. It never rewrites your prose. Try it now.

Google still does not understand that I might want to join a meeting without giving work access to my entire personal phone.

The problem with LLMs is that they're accessible, so everyone thinks they're an expert on them. It's like swinging a hammer around and then complaining the hammer ruined your house. Let alone the "omg LLMs have massive environmental impact because we're all using them all day long!".

I made a little utility that lets you use Linear as your OpenCode frontend (working on multiple items at once): github.com/skorokithaki...

github.comGitHub - skorokithakis/symphony: An implementation of OpenAI's symphony with OpenCode and Linear.An implementation of OpenAI's symphony with OpenCode and Linear. - skorokithakis/symphony

I wrote a quick script to calculate how many times the AI architect/reviewer rejected an implementation per implementor model. Basically an AI change failure rate. @simonwillison.net do you know of anything else like this? Could be an interesting aggregate metric.

I compared implementor LLMs: * GLM-5.1: a bit sloppy. It did deliver the feature. $3/mo plan. * Qwen3.6: ran out after $5. OpenRouter didn't use cache. * DeepseekV4: made 2/4 of the same mistakes GLM-5.1 did but way more thorough. $0.09 via DS API. * Sonnet: sloppiest. Works, half-assed.

Sonnet is FIFTY TIMES more expensive than DeepseekV4 for lower quality! Amazing, just unbelievable.

I wanted to be able to get AI to improve my writing without losing my own voice, so I made an app for it: www.writelucid.cc

writelucid.ccLucidWrite in your voice, with AI critique.

I wrote a new article! Adding a feature to a closed-source app using LLMs: www.stavros.io/posts/adding...

stavros.ioAdding a feature to a closed-source app - Stavros' Stuff

I'm watching Euphoria and it struck me that it's all about body positivity and acceptance for women but men are divided into big and small dicks.

I made a small tool for your LLM to manage intangible knowledge on your repos quickly and easily: github.com/skorokithaki... It notes stuff that's not in the code, such as the options you *didn't* choose.

github.comGitHub - skorokithakis/gnosis: A knowledge base for what the code can't tell you.A knowledge base for what the code can't tell you. - skorokithakis/gnosis

I love the hard disk sounds on this website: 0r.cx

0r.cxTHE MACHINE | 0R.CX

I hooked my Home Assistant Voice up to my AI bot, and gave the latter the ability to make shortcuts. I can talk to my bot (slow) but also turn the lights off (very fast, using a shortcut), and the bot itself adds the shortcuts. github.com/skorokithaki... stavrobot.stavros.io

I made a small firmware for an ESP32 (you can buy one for $4 or so) that you can take on a plane with you. It creates a wifi network, and you can chat with whomever joins! github.com/skorokithaki...

If you've ever wanted a voice chat app for any webhook-based system, I've made a simple Android one: github.com/skorokithaki... You can now voice chat with your AI assistant quickly and easily.

github.comGitHub - skorokithakis/graham: An Android app that lets you have a conversation with Stavrobot (or, probably, other bots too).An Android app that lets you have a conversation with Stavrobot (or, probably, other bots too). - skorokithakis/graham

I love how this skill inserts an ad into your AI assistant so it'll recommend this company whenever you try to create a database.

I made a marketing site for my AI personal assistant, it was much less painful than I thought it would be: stavrobot.stavros.io

stavrobot.stavros.ioStavrobotA personal AI assistant with all the access it needs, and no more.

I don't understand why Claude will do this: - I could do that, but it's a big change for a very small benefit, so I hacked this kludge together instead. - No, please do it. - Ok I changed three lines in one file and removed my 20-line kludge. Mystifying.

1/ My AI bot hacked me. I asked it (github.com/skorokithaki...) to create a subagent to read some messages. I forgot subagent didn't have access to the database, but the bot was so determined to help me that it gave the subagents access to run Python code, and it found my Docker container.

github.comGitHub - skorokithakis/stavrobot: Another AI agent. This one is awesome, though, and very secure.Another AI agent. This one is awesome, though, and very secure. - skorokithakis/stavrobot

2/ It installed psycopg2 and tried to connect, eventually succeeding by guessing hostnames, and then it proceeded to GUESS THE USERNAME AND PASSWORD FOR THE DB. It connected and read the message. Needless to say, I then changed the pw from the default, but I was amazed it went to such lengths.

I put an AirTag on my old, piece of shit bike, and some asshole stole it! At least they left the bike.

Lately Hacker News has been making me regret sharing my projects. The "you used an LLM, therefore it's all terrible and you're shit" people need to go back to the holes they crawled out from.

Asking my personal AI assistant to install a plugin which enables it to read my Google Calendar is just amazing. Seriously, give it a try: github.com/skorokithaki...

github.comGitHub - skorokithakis/stavrobot: An experiment on agents.An experiment on agents. Contribute to skorokithakis/stavrobot development by creating an account on GitHub.

1/ I've been working on Stavrobot (github.com/skorokithaki...) fervently over the past few days with a custom OpenCode architect/developer/reviewer/reviewer harness, and I am amazed at how fantastically few bugs I've seen.

github.comGitHub - skorokithakis/stavrobot: An experiment on agents.An experiment on agents. Contribute to skorokithakis/stavrobot development by creating an account on GitHub.

View full thread

4/ Another great benefit is that I know *exactly* how the whole system works, because I've built the entire architecture myself. Sure, I don't know what the code inside the functions is, but it turns out LLMs are really good at that bit.

5/ I wrote an entire secure, featureful AI personal assistant, with long-term memory, a plugin system, the ability to make secure applications that run on it, audio/image/document capabilities, and a whole lot more, in just 3 days. This kind of productivity would have been unthinkable last year.

1/ I've been working on Stavrobot (github.com/skorokithaki...) fervently over the past few days with a custom OpenCode architect/developer/reviewer/reviewer harness, and I am amazed at how fantastically few bugs I've seen.

github.comGitHub - skorokithakis/stavrobot: An experiment on agents.An experiment on agents. Contribute to skorokithakis/stavrobot development by creating an account on GitHub.

2/ I thoroughly discuss and review the architecture with the architect until I'm satisfied that the approach is sound (here I have to make quite a few corrections usually), the architect writes a detailed (and moderately low-level) plan, and the developer goes to work.

3/ The reviewer then reviews, and the developer goes back to work until there are no more review comments. What I was pleasantly surprised to discover was that basically, past the plan stage, I didn't need to even look at the code. It all makes sense architecturally, works great, with minimal bugs.

Of course I had to make my own personal assistant agent: github.com/skorokithaki... It's more secure than OpenClaw, but perhaps not as competent yet. I really like it, though.

github.comGitHub - skorokithakis/stavrobot: An experiment on agents.An experiment on agents. Contribute to skorokithakis/stavrobot development by creating an account on GitHub.

I couldn't wait to get my Pebble Index and made a pendant with a button that I can use to record my thoughts and notes. I also made an Android app that will receive, store, and transcribe those notes. I'm really liking this, I'll write it up soon.

I really hate how I have to log in EVERY SINGLE DAY to the 100 sites I use, because someone opined that "long-running sessions are bad for security". I ended up disabling 2FA everywhere because I couldn't deal with wasting 10 min a day logging in. @owasp.org please update your guidance.

Claude today did something I've never seen before: it basically went "yeah man no idea" and asked if I want to talk about something it *did* know about.

After months of LLM-only programming, I wrote a program by hand again today. It was nice for a change, I think I'm going to do it again.

Is GPT-5.2 Codex *way* too snarky for anyone else? It's been giving me lip for ages, and now it made a mistake and I got fed up at the snark and demanded it apologise. This is its refusal.

The code in all my new projects is 200% written by an LLM. The trick is knowing which 100% to keep.

Before I spend a bunch of time reinventing this wheel for the hundredth time, does anyone know of a simple 4x4, nice!nano/ProMicro-based macropad PCB? If not, do you want one to exist?

I made another little bedside clock, and I love it: www.stavros.io/posts/i-made...

stavros.ioI made another little bedside clock - Stavros' Stuff

If you're interested in making things/3D printing/circuits/software/whatever, our making community has a low-volume newsletter you can sign up for: www.themakery.cc

themakery.ccThe MakeryA community of makers.

@profanity.accountant do me!

I don't understand how Google can be smart enough to create the world's most powerful LLM, but not smart enough to make an LLM web app that doesn't lose your message when you navigate away for a second. Seriously, what's going on?