AI in music. I've been wanting to post about this one for a while, but I've held off because I know some people don't like AI in music, and I wanted to further evaluate my own thoughts on it. Well, I'm still evaluating, but I thought I'd share where I'm at. I've been writing my own songs since I was a teenager. It's a way of processing my circumstances and the world around me. I've used some form of tech in music since the beginning, though some might not even consider it tech. For example: amps to make an instrument louder than it is, effects pedals to make an instrument sound different from its natural tone, and mastering a final recording to get the levels and compression exactly to my liking. Those are on the music creation side. On the songwriting side, I use Google as a thesaurus to make a line in a verse or chorus work better. All of these came before AI. So far, I haven't used AI to write lyrics because I appreciate the process: the struggle, and finally landing on a line that means a lot to me. On the flip side, I've used AI a lot for the music and melody side of a song. Below are a few songs and how AI was involved. Can't Have Her: suno.com/s/bfNq9TIBhqXQci37 - For this one, I had the lyrics and portions of the melody, and then I had AI fill in the rest. Time was of the essence for this song because I was in the hospital with my wife, and she was in and out of the ICU with everything cancer and chemo do to the body. I played this song every evening at 9 p.m. while walking from the hospital to the hotel. In short, I wrote a song to help me process my exact situation. Looking For Love: suno.com/s/BzRRRbk3OLk0Wkew - This one was written by my dad, Lyndon Bartell. I really like the song but wanted to try a new musical style and improve the production. Dad passed away a few years ago. He was a prolific songwriter, and I hope to continue "remixing" his songs. It's a great way to remember a loved one. Peace And Rest: suno.com/s/8nhDsuqA8NCFrR5K - I wrote the lyrics and melody for this one but didn't want to do a formal recording, since my voice isn't exactly easy on the ears. Instead, I sang the melody into my Mac's mic, uploaded it to Suno, told it the style I wanted, and voilà: a song that sounds much better than if I had done it myself. And then there are what I believe to be 100% AI-generated songs, like this one from Bleeding Verse, "If You Ever Hear This": open.spotify.com/track/5FQ1p…. My question about this song is: do I care that it was AI-generated? No. I care that I enjoy listening to it. That said, it doesn't bother me that others don't like it simply because it's AI-generated. I understand the appreciation for something authentically human. Where do you land on AI and music? Pic for attention. AI generated ;-)

Businesses, your websites need to allow bots! I had a 10-minute drive from my house to an area with several hardware stores. I needed a 23mm wrench for a lower control A-arm bolt. I put @bot to work and told it to get me inventory and pricing from two stores, @Menards and @HarborFreight. Menards' website blocked the bot. Harbor Freight allowed it. Harbor Freight got my business yesterday.

How did your representatives vote on the issues that matter to you, from Congress down to your city council and school board? Most of us have no idea, and digging through bills and meeting minutes is nobody's idea of a fun evening. The idea came from a good friend of mine, a political science professor, who wanted an easier way to learn this information. So I ran an experiment. Working with Grok Bot, which did all the research, I built Legisun in about three hours total. It's a free civic site (a nod to Legistorm). Enter a street address and you'll see how your federal, state, and local officials voted, searchable by topic or keyword. What's under the hood: • Real data from Congress.gov, senate.gov, and LegiScan for the Minnesota legislature • City council, county board, and school board minutes across south central Minnesota (Mankato, North Mankato, New Ulm, St. Peter, several counties, ISD 77, and Maple River), about 1,300 local motions so far • A short summary for every bill and motion: the official CRS summary when one exists, otherwise a clearly labeled AI summary • Links back to the source bill or minutes • Daily refreshes that prioritize the boards people look up but that don't have data yet • Hosted free on Vercel For now it only covers south central Minnesota. Want to kick the tires? Try the Mankato post office address, 220 E Main St, Mankato, MN 56001, at legisun.vercel.app/ How are you using AI for the midterm elections? Pic for attention. This is my lovely wife, pregnant with our fifth and final child 17 years ago. She's thrilled to have a bun in the oven. We road-tripped to Chicago for the Rock 'n' Roll Half Marathon. I had a lot more hair then. Not wanting to miss an opportunity to fit in, I took out all the stoppers as I strutted on Chicago's streets.

New daily driver: I've switched my M1 Mac laptop over to Omarchy, and the AI-first approach has me hooked. Want slower scrolling? I type "slow down scroll speed" into Claude and it's done. Immediately. No restarting anything. Key remapping works the same way. I told it "change Cmd+Tab to rotate between apps in a group." Done. Even installing software that's "not yet supported on Omarchy" is far simpler than it used to be. If a problem comes up, Claude reads the error, looks at your system, and gives you an expert fix. It's worked every time so far. Pic for attention. This is one workspace with several apps grouped together (Claude, X, Grok Bot, VS Code), and I can Cmd+Tab through them just like on a traditional Mac. Cmd+1, Cmd+2, and so on jump between workspaces.

One promise I've heard about AI, and about the technologies before it, is that it will give us time back in our lives. I've learned that, with AI as with earlier technologies, that has never been the case. Without discretion and discipline, we tend to fill the empty spaces right back up with more busyness. I'm very guilty of that. When I go on vacation, it's usually not until the third day that I finally get comfortable enough to stop looking for something to do or accomplish every five minutes. My mind is used to racing. So I suppose that means a guy needs to take the bull by the horns, grab moments of serenity, and let them be filled with nothing but pondering. Pic for attention. Remember the Sony Discman? I don't know what my son got into earlier in the day, but it looks like he had fun. Unfortunately, bath day wasn't until next week.

Maintaining an app: tech debt is real, and one part of it is keeping your libraries up to date. The other day I set up Dependabot in my GitHub CI, and it surfaced 17 pull requests for things that needed updating. Bleh. I went and made a second cup of coffee, thinking I'd need to work through each of them. I honestly don't know why that was my first thought, because I've been pushing Claude hard lately. So instead, I pasted the PR list into Claude and said, "Fix these." I didn't tell it, "Fix the easy ones and talk with me about the hard ones." It just did the right thing. For patch and minor versions, it pulled the code to my localhost, upgraded, ran the tests, and merged the PR. For major version upgrades, it did the same but stopped before merging to main and asked for my permission. I'm thankful Claude is getting to this point, protecting me from myself. I didn't tell it to be careful; it just is, most of the time :-) Pic for attention. These are my two youngest boys in June 2020, at the beginning of COVID. I gave them full access to all my power tools and boxes of screws and nails, and let them build whatever they could dream up. That said, this was the before pic. The after pic had a lot of blood. But boy, did they have fun!

Yesteryear, it was a challenge to justify a single software license that cost a few hundred dollars a year, usually for software I'd used at a previous employer and knew would deliver a solid ROI. I don't know exactly how much my employer spends on AI for me as an individual, but it's significantly more than that. Last month, I personally consumed well over $2,000 in AI tokens. My, how things have changed. I like it! Now, that would be a waste of money if the investment didn't produce results. What's interesting to me is that I can justify the expense within the same day, let alone the same week or month, because of how much more work I'm able to accomplish. To that point, if you don't have a chart showing year-over-year productivity, you need to make one. I have charts that compare total stories and points completed year over year, and it's fun to see the "AI difference." I also have a 90-day running graph that shows how much I'm improving each month. Pic for attention. I've dragged all five of my kids to the hardware store for projects over the years. Evidently Elliot grew tired of me traipsing through the aisles and decided to take some downtime of his own. Check out those socks!

Mad skills. I was chatting with another geek and mentioned a Claude Skill I'd developed. Later he messaged asking how I authored it. Here's my response. Yes, I asked my own questions. I'm never alone — all the voices in my head keep asking anyway ;-) Yep, I use Claude Skills extensively. In short, Skills are just text files with instructions, at least in their simplest form. Why create one? Glad you asked. Whenever you're sick and tired of manual clicks and think "I'd rather be in the woods," that's when you create a skill. Example: I have one named /triage. When invoked it: • Queries production logs • Searches Jira for existing stories • Queries Slack for related conversations • Reviews my local repos to locate the code issue How do you create one? Prompt Claude like: "I want to create a skill named triage. I want it to search logs at this location. I want it to search Slack and Jira for similar issues. Look at my local repos and try to identify if there's a code issue." Claude stores it on your machine, and you can use it from any Claude Code surface (terminal, VS Code extension, desktop). Usage looks like: /triage customer support is reporting that the API endpoint POST /customers is throwing 500 errors. Can you help a guy out? Pic for attention. This took mad skills to get the mower up on this rock. It was accomplished by one of my offspring. Those Snapper riders were bullet proof.

Claude, can you mow the lawn? I've said this before, but I'm constantly amazed at how I don't dream big enough with AI. I keep trying to do things the way I've done them for 28 years as a software engineer. "Think of Claude as a co-worker." I believe I heard that from Boris (find him on x.com). Ok, so a co-worker. As with any work relationship you're onboarding, you need a feel for: what can you do, and how much supervision do you require? Claude is no different. Same principle for tool access: if you don't give Claude the tools to do the job, you'll forever be tied to the processes you're trying to hand off. You MUST eventually give Claude access to your work tools. For me that's GitLab, Jira, Slack, Outlook, source on my machine, Grafana (logs), etc. Once you grant that access, you start seeing how much you can actually delegate. Example: I created a Claude Skill named /triage. Run it with a screenshot from Customer Support and it queries prod logs, Slack, and Jira, then summarizes. Those steps used to take me 45–60 minutes. Now I prompt for ~30 seconds and go do something else while it runs (~5–10 minutes). Pic for attention: my oldest son, Ethan Bartell. If Ethan were Claude, this is how you offload 100+ lbs of pack so a guy like me can make the 3mi hike. Superior Hiking Trail last month — worth it if you want awesome trails.

Invite critique. I do a lot of the deploys for our app. The other day another engineer did the deploy from the docs — mostly written by me. There are a ton of steps: creating stories, setting dates, assigning things, attesting that certain rules have been adopted, creating Git tags in multiple repos… you get the picture. At the end he was thoroughly frustrated that it took so many steps, and said so in our retro. That afternoon I spent a couple of hours authoring a /deploy skill that automates ~90% of the mundane steps. Here's the thing: if he hadn't done the deploy, we probably would have kept going the manual way — because I could do it faster from memory. What in your daily / weekly / monthly routines needs some AI automation? Maybe you're like me and have a process with plain old bad, complicated documentation. Time to let AI have its day on that piece of work. Pic for attention. It's Gustie football season and that means Friday linebacker dinners! 45 brats and 24 hotdogs were queued up on the grill for this meal - linebackers can really pack it away! This year's meals hold special meaning because last year Amanda (wife) and I were in the hospital for the majority of the year getting chemo. Thank you Jesus that we get to host these young men and be a sliver of their lives.

"Fundamentally curious." A stranger said this in an exchange on X, describing what tomorrow's worker will need in light of AI. I agree. The term "generalist" is also being thrown around a lot, and I agree that this personality and role will matter a great deal moving forward. Why? Because both mean conquering the unknown and being willing to wear many hats. That will define the next few years of AI, if not longer. Some engineers will step into this new role. But even more, we'll see people who were always waiting on other departments (including engineering) to deliver what they knew was right for their customers. Those people are steadily becoming unbridled. Honestly, it's fun to watch. Be fundamentally curious. Pic for attention. This is Elliot in his early days, when he'd escape his crib by climbing onto his dresser. It took my wife and me a hot minute to figure out how he was getting out. Eventually we captured his "foundational curiosity" for escape on camera. :-)

"AI is causing us to lose critical thinking." I've heard this a number of times over the past few months. After pondering, some thoughts: Early in my career I wrote C — bare-bones, lots of work, deeper thinking. Around the same time I was using RPG (IBM), which didn't require pointers and let me operate at a higher conceptual level. Then Java, PHP, and eventually the gold mine: Ruby on Rails. Easiest of the bunch. Each language/framework was designed to ease implementing business rules. Each one required less critical thinking about *how* to implement the need. Enter AI. Some say it removes the need to think critically. I get why — students using it to write papers, etc. What most don't realize: future employers (and current ones) will frown on people who can't think beyond AI for emails, marketing copy, decks, reports, data entry, and automations from scratch. Employers will demand bigger, more complex problem-solving. The next generation needs critical thinking *and* pioneer-level curiosity and experimentation. Are we training for that in high school and college? Most colleges are reactive to the job market — and rightly so. Who wants to pay for a degree that *might* have prospects. I'm curious whether we've ever seen a job shift this big, this fast. Every keyboard job is up for debate. Every keyboard job can likely automate a large % of its tasks. Are you an automator pioneer who's curious enough to experiment? Pic for attention. This is our now adult Elliot who, aiming to look like a scientist for science fair had us get him some non-prescription "scientist" glasses. Oh to live that week again.

Does your tech stack still matter? At Rails World last week, DHH, the creator of Rails, said 37signals has gone "pencils down" on hand-written code. Agents write the implementation; engineers direct and review. And HEY is being rebuilt as native apps on a Rust backend. The creator of the framework built around developer happiness is now choosing tools based on what runs best, not what's nicest to type. After 28 years of writing software, my take: the stack still matters, just for different reasons. Matters less: • How pleasant a language is to write by hand • Whether you personally "know" the framework • Clever abstractions built to save keystrokes Matters more: • Strong conventions an agent can follow • Performance and hosting cost, now that code is cheap to produce • How quickly you can review and verify what gets generated Even 37signals learned the hard way. A pile of reasonable-looking AI pull requests left Basecamp's architecture "like Swiss cheese." So yes, reconsider your tools. But the question is no longer "what do I enjoy writing?" It's "what can I trust, verify, and run cheaply when I'm not the one writing it?" Watch the whole keynote here: youtube.com/watch?v=vDjW_dRy… Pic for attention. Beef steak for supper is always the first meal on our camping trips. I am now hungry :-) What will your stack look like a year from now?

Idea for where you should push your AI agent hard: Anywhere it's easy to undo if something goes wrong. For me, that's basically anything that doesn't involve direct customer communication. Specifically — there's every reason to use AI with your task system (Jira) and say things like: "Create a story about xyz, assign it to me, set the scrum team to xyz, set priority to high, and start writing the code." Or: "Respond to the pull request comment with the findings from this triage session. Be succinct." Or: "Post a response to the Slack message this originated from declaring the results from running the /triage skill. Mention Tamra, Vitor, and Tim." Early on with AI I was incredibly picky about verbosity and how it wrote code. I'm realizing opinions matter a lot less than business progress. Claude is incrementally becoming my command center — the place I direct from. What tips do you have for how you're using your AI agent? Pic for attention. My oldest son, Ethan Bartell, is on the Gustavus college football team. He's a captain this year and was in charge of pulling off a captain's practice meal. This is the pork loin I volunteered to cook up (this was only half of the batch). Feeding 40 football players is something amazing to watch. Even though we had 20 lbs of pulled pork they still needed to buy another 50 hotdogs. Everything was gone in the end. If you ever find yourself cooking for football players plan for 1.5 lbs of meat per player. :-)

"I am going to be paying attention to how long you're gone." Overheard from a parent talking to their child after the 4-year-old asked to go inside to get a drink of water. Such good parenting — and it's funny. It made me think about how long I wait for Claude to respond. That's why I end up with three to seven prompts going at the same time. I noticed something this week. For years I've had to context-switch dozens of times a day — and not minor ones. Sales call → coding → project planning → architecting → triaging customer issues → back to coding. I'm finding that comfort with frequent context switching is a superpower when working with AI agents. You basically have a team around you that you direct. They come back with questions or finish the work. I'm incrementally making my AI desktop client the "window" into every other app in the enterprise (more on that later). Frequent long-term context switching eventually burns people out. Right now it's fun. In a couple years — will it be? Time will tell. But burnout has always been part of tech, hasn't it? I've had the equivalent of a pager for ~25 years. I'm almost always available and almost always "on." It's been so long I don't really know another way. I talk with my wife about finding my "retirement job" in the next 10 years so I can see what a non-tech life feels like. I actually desire manual labor of some sort. AI can't give me farm-to-table corn with my grandma. That part still has to be human. Pic for attention. This is me and my grandma. She's taught me a lot over the years, both from observation and putting things into practice. Here we're engaging in the multi-generational practice of harvesting corn, shucking corn, cutting the kernels off the cob, and finally, putting it in freezer bags and the freezer. Closest thing to farm-to-table I've experienced, and I enjoy every minute of it because it's interacting with a person I love. AI can't give me that.

Of Trips and Google Maps. As I was "AI Coding" recently, I realized something. Because I can produce and test-drive new features locally so easily, I don't need thorough plans. I AI-code a feature, play with it for a day or an hour, then either remove it or refine it. Even better: I excitedly show my wife and kids the latest feature, and the looks on their faces tell me everything — instant feedback from people who live on TikTok / Snapchat / etc., platforms I'm not entrenched in. This reminded me that I no longer plan road trips on a map. I just trust Google Maps to get me there. And if I miss a turn because I'm daydreaming, I get rerouted and try again. AI seems similar — at least if I let it. Pic for attention. That's my water gravity filtration system for camping last week. Had to go with a contingency destination given the BWCA was burning at the time of my trip.

AI selected my vendors. I told Claude to create an email verification feature for new accounts. It implemented the feature well — but what really surprised me is that it selected a vendor as part of the implementation. I hadn't given it info about my mail server. Instead it wired up resend.com so verification email could go out to new account holders. I had no idea this vendor existed. Claude automagically put it in my app and made the decision for me. I'm still processing what that means. Btw, Resend is working great. What "AI surprise" have you had lately? Pic for attention. This is me about 9 years ago when I sold the software I wrote. I don't miss traveling but I sure do miss the people I met in my travels.

How to make my Claude /code-review skill better? · ShareAI shareai.fyi/posts/90776989-d…

Four Claude sessions running like junior devs. The pic is the original crew + a Raspberry Pi.

Code in my sleep. I had a new first. I was tired and had a large task for Claude. I'd built an iOS app and wanted the same for Android. I turned on Claude's Auto mode (stops the constant prompting) and slid it up to Ultracode. Let's burn some tokens, baby! Then I took an hour nap on my deck under an umbrella. Why does it matter that I slept under an umbrella? Who knows! :-) When I woke up, the Android equivalent was complete. What's more interesting: Claude did it without the Android SDK installed. So then I had Claude install the SDK and install the app on an old Android phone. It worked. Mind blown. Aesthetics needed attention, but the app worked after a single prompt. If I had to guess, I saved at least 30 days of coding. It took me longer to set up OAuth in Google Cloud Console than to write the Android app. Pic for attention. A hobby I've had for a while is sewing my own bridge hammocks. Best sleep you'll have in the woods, and the woods is the best place to sleep. Location is BWCA South Kawishiwi River.