Github link: github.com/bluestemso/n...
Blog
Total Records: 80
Finally got around to building a little tool I've wanted for the longest time. I'm colorblind, and all of the color ID tools I've found are for designers, so they return a hex or RGB, not a name. I just wanna know if it's blue or purple, man.
Questions are free now. Been difficult for me for two reasons: 1. Years of encoding reminding me to not waste others' time with questions I could answer on my own, 2. Despite tons of rich answers, I can only absorb so much in a day. Answers far outpacing knowledge gain.
Absolutely loving MiniMax M2.7 through OpenCode Go. Feels like flying.
The issue with the memory/meta-frameworks like GSD and Superpowers is that when I enter one out-of-framework prompt, it takes 2-3 prompts to get everything back on the rails. Gotta figure out how to push all prompts through without blocking outside functionality.
There's a strange second-order effect from having (effectively) unlimited access to tokens. I didn't realize how much mental energy I was spending deciding if I should open a session because I was planning out my token budget for the day/week.
Now I switched over to open source models, and I'm moving so much faster because I just kick off the session and keep moving.
Very curious to see what kind of UI we land on for nested agent workflows.
Trying out the open source models on OpenCode Zen and realizing just how overkill 5.3 Codex is for many tasks. The speed from the OS Zen models really improves the experience.
Switched to GSD from Superpowers, and boy oh boy what a difference. Walking through the decisions one by one goes so far in terms of keeping my mental model of the project up-to-date. Highly recommend for newer devs.
What's the correct way to fix a subagent hang in Opencode? Is there some way to just kill the subagent and restart? Or need to kill the entire session and prompt with "continue"?
One day devs will be measured by token efficiency, I'm sure of it. I spend a ton more tokens than is necessary simply because I spend a lot of time exploring the possiblities and codebase.
Do we have a name yet for the GSD, gstack, superpowers kind of tools yet? I feel like GSD's description of "meta-prompting, context engineering and spec-driven development system" is accurate, but just doesn't quite roll off the tongue.
Antigravity monthly credit system just seems way off in today's market. I'm regularly checking my Codex/Claude session/weekly usage limits, already planning what I'm going to give them when the next session starts. It's day two and I'm out of Google credits, so, uh, guess I'll see you next month?
I feel like such an in-betweener right now. To experienced devs, I'm a vibe-coding larper. To non-devs, I'm a nerd maniacally ranting about the AI tsunami heading towards land. Just can't win with these people.
Even two years into learning development, I'm astounded by the massive number of decisions that need to be made in order to produce even a simple app.
Seems like lately the GPTs suggesting more and more 3rd party tools. I ask for a relatively simple local script and it suggests 2-3 libraries to achieve what a 200 line script should be able to do. Wondering where that's coming from.
Google's Stitch feels directionally right in a lot of ways. Still clunky as hell, but definitely an interface that's >>> than text to a coding agent.
Where's the suggested folder feature in Google Drive? Seems like such an easy win. When I add a new scanned receipt to my home folder, Drive offers to make a podcast about it, but still can't figure out which folder it belongs in???
NB2 š¤
With everyone doing speech-to-text now, how long do we think until the first "dev helmet" comes out that enables unheard of dev density on a per sq ft basis?
No but for real though, curious about how tech spaces are going to change. Will it eventually just be normal for everyone to be talking to their machines all the time?
Even with STT in messaging apps, I often type instead just because I'm self-conscious about others hearing some out of context garbage coming out of my mouth.
Set my OpenClaw up on an old laptop, gave it a dedicated Obsidian vault that's synced to my machine. So much better than trying to read large blocks of markdown in a chat window.
Also wrote a script that scans my personal Obsidian vault (which my claw does not have access to), and shares anything with a certain tag or in a certain folder. This way I can share notes with it without having to modify those notes or make copies. The script also handles un-sharing.
I feel like I'm missing something, but managing my skill library with OpenCode (or Claude or Codex, all have the same issue) isn't intuitive at all. Trying to build a library where I can easily toggle skills on and off per project and even per-session and not have to open up a config file each time.
From CRUD monkey to API key monkey. Look how far we've come.
Just discovered interactive mode when listening to audio overviews in NotebookLM. Hard to imagine a world in which that doesn't become one of the most popular ways to learn about new topics, especially in academics.
Driving a lot of AI tools feels like driving the Suzuki Escudo Pikes Peak Special in Gran Turismo. If you're not vigilant, everything just spins out of control real fast.
Spending 3 hours setting up an automation to avoid a total of 15 minutes worth of work over the next 6 months == peak me behavior
Inspired by @simonwillison.net and his tools page, decided to start publishing little HTML/CSS/Vanilla JS "gadgets" on my personal site: gadgets.taylorlearns.com
I'm doing my best to learn tmux, so started out with a little dojo to hopefully burn those keyboard shortcuts into muscle memory. gadgets.taylorlearns.com/tmux-dojo/
Amazon's coding agents just don't quit
First, by learning what the agents were and werenāt good at, the team could switch from babysitting every small task to directing agents toward broad, goal-driven outcomes. Second, the velocity of our teams was tied to how many agentic tasks they could run simultaneously. Third, the longer the agents could operate on their own, the better.
When did the amount of time a coding agent can run without intervention become a priority metric? I understand how it can be useful and that we need agents that have the right ability to overcome obstacles, but it feels like a lot of the labs & agent makers out there are focusing on it as it's own metric, which I don't think benefits users nearly as much as it benefits those labs & agent makers. And I think this quote kind of exposes that. They have rationale for points one and two, and then agents running for longer is just... 'better.' Feels like a setup for one day these agents taking 10 hours and 1 million tokens just to come back and say that they updated a single CSS class and now your problem is fixed.
Seems like there's good use for a kind of 'soft-allow' permission with coding agents. When pair-programming, I enable `git commit` permissions but I have instructions not to use it unless I say so. Worked well so far.
Tuning 3D-printed flutes with a vibe-coded tool
I love what this guy's doing.
He's a purveyor of 3D-printed flutes, both the flutes themselves but also the 3D printing files so you can make them yourself. Cool enough in its own right.
But what really caught my attention was that he figured out how to vibe code his own tuning software. So he prints a flute, then runs the software while he plays the flute and gets immediate feedback on how well it's tuned. He then uses that data to modify the design and just keeps iterating.
I haven't spoken to the guy, but based on some of his reels, I'm assuming he never could have done this without AI-assisted development.
Now I'm off to print some flutes.
#3D printing
#vibe-coding
Unexpectedly, a deer briefly entered the family room
The animal mix-up is not ideal, but itās not a major hurdle to usability. I didnāt seriously entertain the possibility that a deer had wandered into the house, and itās a little funny the way the daily report continues to express amazement that wildlife is invading. Itās a pretty harmless screw-up.
āOverall identification accuracy depends on several factors, including the visual details available in the camera clip for Gemini to process,ā explains a Google spokesperson. āAs a large language model, Gemini can sometimes make inferential mistakes, which leads to these misidentifications, such as confusing your dog with a cat or deer.ā
Google also says that you can tune the AI by correcting it when it screws up. This works sometimes, but the system still doesnāt truly understand anythingāthatās beyond the capabilities of a generative AI model. After telling Gemini that itās seeing dogs rather than deer, it sees wildlife less often. However, it doesnāt seem to trust me all the time, causing it to report the appearance of a deer that is āprobablyā just a dog.
I can't tell you how amusing I find it when the AI doubts the user.
There's just something about self-hosting
Self-hosting just hits different.
#Hetzner
#Open WebUI
#OpenRouter
#VPS
#self-hosting
been looking for interesting data sets to practice data analysis and using llm's for analysis, so i used google takeout to grab my google search and acitivity data. large enough to be a challenge and super interesting. highly recommend.
Would love to see more open benchmarks like this for image editing, and other uses of LLM tools as well:
genai-showdown.specr.net/image-editing
via @simonwillison.net
Frustrated by the lack of testing tools in no-code platforms. I spend a solid 30-40% of my total time on a no-code build doing manual testing that could easily be automated on a project built with code. Who's doing testing automation for no-code platforms?
When using a chatbot interface, why in the world is it the standard practice to have the window auto scroll as the response is generated? who thinks people read from the end of a response to the beginning? unless you're using a slow local model, no one reads as fast as tokens generate.
Big 'first' today. Started scoping a no-code build for a client and thought to myself, "This would actually be easier to build with code." Progress!
Reading Andrej Karpathy's animals vs ghosts blog post (v interesting in its own right), and love that he included a link at the end of it to a ChatGPT conversation that's pre-loaded with his blog post as well as all of the context from the post (he references a podcast and an essay). So helpful.
If I had a nickel for every time a coding agent opened up a shell into a dev container and then just sat there...
we gotta figure out how to teach these llm's that it's okay not to know, you know?
Check out PokPok! None of the ad BS, very well designed, the mini games change over time, and most importantly, my kids love it.
But they do make you pay for all those privileges, and dearly. I've also found that all of the official apps from PBS avoid this garbage behavior.
Such a great way to use text classification.
git.sr.ht/~kqr/tagnostic
Realizing just how much 'context engineering' applies to my own brain
So much of learning to build software for me has been understanding what I can safely ignore in a process. Guides, libraries, platforms, etc are just so rich with optional stuff, and I've always wanted to understand it all, but then never get anything done. Focus is so crucial.
This is the kind of thing that vibe coding is making possible: flotilla-orpin.vercel.app
It's time we all got a bit more creative when writing demo projects for tutorials. Looking at the demo for using docker with django, and they named the sample project 'django-docker'. If you're a newcomer, `docker build -t docker-django .` is confusing as hell. Call it 'my-web-app' at the least..
Great writeup on the value of an open social protocol. Lately wishing I didn't have to pick between the people I want to follow and the platform they're on, so here's to hoping the open social timeline accelerates!
Looking for an open-source implementation of Duolingo Video Call, or at least a good system prompt to achieve a similar kind of flow.
If you're looking to vibe code a clone of some popular app, it's much easier to just find an open source version of it (there are plenty for just about everything out there), fork that version, and then just make your tweaks. Save yourself some tokens and trouble!
I know I'm dreaming here, but trying to imagine the world in which writers, scientists, developers, artists, etc are excited about contributing their works to an AI model. A world where access to that model is free for everyone and no one profits from it, it's just the people's model.
Window management in Mac OS is just wild to me. Although just discovered Settings > Desktop & Dock > Mission Control (Group windows by application) and I'm amazed that isn't default.
Blocked Apple account
I paid my $100 to become an Apple developer, only to be blocked 24 hours later. And Apple couldn't care less.
#apple
#mobile
Just one blog
My journey navigating the sea of options for publishing my blog and portfolio site.
#blogging
#django
Seems inevitable that we will have fine tuned models for each language/framework combination. Kinda surprised I haven't run into one yet.
I released my first production app to the Google Play store today!
The app is a timer/reminder for resting your eyes when working at a screen for hours on end. Itās intentionally simple as the purpose of making it was just to go through the process of testing, packaging, and releasing an app to the store. I didnāt want to include any kind of payment processing or user data so as to minimize security risk.
But now thatās done, time to move on to the next one.
#android
#mobile
I know full well the right way to improve as a dev is to build every day.
The real obstacle for me is deciding what to build. Ideally, I would be working on a unique idea that at least has the potential to become a valuable product. But boy oh boy am I great at talking myself out of just about every single idea. So absent that, IĀ shouldĀ at least be building replicas (another to-do list app, recipe manager, etc.) just to get in the reps, but then I convince myself that is a waste of time because none of them would ever be able to stand out in the crowd.
Gotta get out of this loop. Just gotta build.
#misc
Time to get this blogging party started. Iāve heard consistency is key on these things, so thatās what weāre going for.
#misc
You mention around the 8 minute mark that there are some newer non-profit newsrooms that are doing really well. Any chance you might be willing to share a few of those? Always on the hunt for journalism that isn't horrible.
Thanks!
Setting up and using ssh keys has always given me way more trouble than it should
From a cost standpoint, seems like the tradeoff on web app v. mobile is that with web apps, server costs are significant but you get to keep nearly everything you make, while with mobile apps, server costs are insignificant but Google/Apple take a big cut. That about it?
Made a remarkable amount of progress on my first LLM-powered feature today. Slowed way down and planned more. Steady is fast and all that.
My loop seems to be spending 2-3 days bashing my head against all the internet/AI content about a topic, followed by a brief moment of clarity in which everything clicks. Today is a head bashing day unfortunately.
Plowed through some tough Firebase authentication API stuff this morning and felt great about it. Then spent the past hour wrestling with the setup of a Google Group...
Aaand turns out the problem wasn't with Google Groups, but entirely with me. This is why I don't post on socials.
I've been an Excel guy for years, and so when I first encountered the idea of a NoSQL database I was v confused. Then I heard "we're optimizing for reads, not writes," and immediately understood.
Aaand forgot the link:
youtu.be/JhCl-GeT4jw?...
First saw this video a while back and haven't been able to stop thinking about it. Strange to think about how many practices and artifacts in development exist for human benefit, not for the benefit of the program per se. Then thinking about which of those are still useful when AI is developing...
There has to be a reason that GitHub doesn't natively support downloading a subdirectory of a repo, right? Right?!
I feel like I'm probably doing something wrong, but man I spend an awful lot of time wrestling with run/debug in Flutter. Django local server seems so much more reliable.
Also haven't been able to get the Android emulator running on my Windows machine. I suspect it's because I have WSL occupying about half of my 16 GB of memory, so not enough to run the emulator on either Windows or WSL. But I refuse to believe that I can't both have WSL and run an emulator...
Careering by Teal Process & Company
When individuals āchange careersā or ābuild a new careerā theyāre careering! Theyāre swerving to another lane, jumping over the high credential barrier to do so. This introduces the The Careering Model of working. Its based on the verb of career, meaning to move swiftly and in an uncontrolled manner in a specified direction.
Such a carefully considered and well thought out approach to ācareering.ā Absolutely love this.
#Career
Iām nowhere near qualified yet to comment on the actual merits of Tailwind as it relates to most of the other common CSS frameworks (Bootstrap, Bulma, etc), but I do know why I chose to implement it on this blog.
I donāt even remember where I heard or read this, but an experienced dev said that he could immediately identify a site that used one of the more opinionated frameworks, whereas Tailwind wasnāt nearly as easy to spot.
I decided in that moment that using anything other than Tailwind would be the equivalent of showing up to prom in what I thought was the coolest, most unique tux, only to discover that two of my least favorite people were wearing the same thing.
Now itās been a week since I implemented it (with considerable help from Cursor), and a one hour struggle with a simple alignment issue has shown me just how much Iām in over my head in the CSS waters.
However, and perhaps misguidedly, Iāve decided that Iām not giving up and Iām riding the horse I came in on.Dave Gray's Tailwind CSS Full Course for BeginnersĀ has been great so far, although even it is moving pretty quickly for me. And Iāve already gone through most of the CSS stuff on freeCodeCamp.
Yet another example of how AI code assistants can quickly get you over your skis as a novice. Another a reminder to start with basics (even if it makes me look... basic) and then upgrade later as I get more proficient.
#CSS
#Frameworks
#Tailwind
Just finished three of the bash tutorials on freeCodeCamp and feels like Iām driving on a freshly plowed road. Not sure how universal this is, but Iāve been blindly copying so many bash commands along this journey so far, and it bothered me so much that I didnāt understand why so many of them worked, or how they worked. I realize that it might be putting the cart in front of the horse a little bit, but I would definitely recommend learning bash earlier than later. Between having the context while youāre developing and also being able to significantly improve your development efficiency, itās such a great tool to have on your belt.
#bash