Brandon Davis

Amtrak and Skiing Whitefish, MT

Wed, 18 Feb 2026 20:00:00 +0000

48 hours on the empire builder and 2 days in Whitefish.

Claude Code SDK is pretty wild

Wed, 11 Feb 2026 22:00:00 +0000

I’ve been using claude code on and off since it launched a year ago. Since the Christmas holiday, it seems a ton of folks outside the software industry have discovered it as well.

Despite its popularity, I’ve seen almost no discussion of Claude Code SDK. It’s essentially a typescript or python wrapper to the agent harness that allows you to fully manage the handoff between agentic and procedural parts of an application.

Beyond `--print` mode

If you’ve discovered the --print flag, you’ve almost certainly scripted up some terrible nonsense to run it as a subprocess.

Stop that right now!

The SDK documentation is dense. It does not take the time to explain the basics of the turn loop or tool hooks. I suspect that most people who have attempted to use the SDK have done so in a one-shot sort of way, by giving a prompt and letting Claude cook until it yields.

Control handoff and tool use

Claude code is just a while loop.

until llm says stop:
 give llm some inputs
 do whatever it says (read file, run bash)

When you build an agent SKILL.md, you hand-wavishly describe how to use an API or function or script. Claude reads those instructions and then uses its existing tools (read/write/bash/grep) to do whatever the skill describes.

Claude is like the little dude inside the big dude’s head

– Will Smith

When you build custom tools with the SDK, you must implement an MCP tool interface, run it synchronously in-process, and hand the output back for the next turn.

A Toy Example

The official docs provide this example. It’s a bit unsatisfying because it hides the details, so let’s reimplement without the automagical tool_runner.

from anthropic import Anthropic
import json

client = Anthropic()
max_turns = 10
SYSTEM_PROMPT = "System Prompt Here"
MESSAGES = [{"role": "user", "content": [{"type": "text", "text": "Tell me the weather in San Francisco." }]}]

# This is the MCP spec!
TOOLS = [{
 "name": "get_weather",
 "description": "Get the current weather in a given location",
 "input_schema": {
 "type": "object",
 "properties": {
 "location": {
 "type": "string",
 "description": "The city and state, e.g. San Francisco, CA"
 }
 },
 "required": ["location"]
 }
}]

def get_weather(location: str) -> str:
 return json.dumps(
 {
 "location": location,
 "temperature": "68°F",
 "condition": "Sunny",
 }
 )

for _ in range(max_turns):
 response = client.messages.create(
 model="claude-sonnet-4-5",
 max_tokens=4096,
 system=SYSTEM_PROMPT,
 tools=TOOLS,
 messages=MESSAGES,
 )

 if response.stop_reason == "tool_use":
 tool_results = []
 assistant_content = []

 for block in response.content:
 assistant_content.append(block)

 # Do what claude says
 if block.type == "tool_use":

 if block.name == "get_weather":
 result = get_weather(block.input.get("location"))
 tool_results.append({
 "type": "tool_result",
 "tool_use_id": block.id,
 "content": json.dumps(result),
 })

 # Append the request and response to the session log:
 # Claude says "Please run this tool"
 messages.append({"role": "assistant", "content": assistant_content})
 # User says "Here you go"
 messages.append({"role": "user", "content": tool_results})

 elif response.stop_reason == "end_turn":
 print("Claude exited.")

You’d miss it in the beta.tool_runner wrapper, but tool use blocks append to an ever-growing messages array that gets fed back into the client every turn.

Takeaway

Using the same harness framework and a completely different set of tools, claude code can do tasks entirely unrelated to writing code.

Renee Good was killed by ICE agents in Minneapolis yesterday

Thu, 08 Jan 2026 22:00:00 +0000

Editorial note This was unpublished temporarily while I had a think about whether what I wrote is spiritually constructive. I decided to put it back.

I live in Minneapolis, where on Jan 7 Renee Good was killed by an ICE officer while trying to flee from another officer attempting to force her car door open. Moments before, she is seen on camera waving out her window for the ICE vehicle to pass using the open lane in front of her.

Neither the footage nor a single eyewitness account agrees with the DHS official statement. All bystanders perceived that she was scared and trying to escape.

Now, the FBI have forced MN state authorities out of the case

The reality

There are many different angles. Here they are in a Google Drive folder. Watch with caution.

Kristi Noem said at a conference that

As they were attempting to push the vehicle [out of the snow], she said a woman “attacked them and those surrounding them” and “attempted to run them over and ram them with her vehicle.”

These videos show no such thing. The officers get out of a running vehicle, and could have chosen to drive by without stopping to confront her at all. Turn the sound on and you’ll hear how alarmed bystanders are by the sudden escalation: nobody on either side of the conflict had clocked Renee as a threat or someone who was behaving erratically.

Other information

A final note

If this feels out of place or uncomforable, that is sort of the point. It is not possible to compartmentalize between our professional selves, our political selves, and who we are in our comunities and families.

I want these people out of my home.

Thoughts for 2026

Sun, 04 Jan 2026 15:30:00 +0000

I celebrated the new year by having covid and coughing a lot. It caused us to cancel a lot of plans with friends and family and cut holiday travel short.

2025 was good to me, though.

I married my wonderful partner of 10 years in June. The joy and gratitude I feel overwhelms everything else that happend this year.
My favorite place to be is still on a bicycle in the woods. The consistency of this fact is comforting.
I want to build a writing habit. Got a jump start on this last November, but I’d be satisfied with producing 2-3 short posts a month consistently.
I over-indexed this past year on trying to solve my device addiction by selectively cutting out specific websites. Live and learn. I’ll try something else and try to be patient about it.

My favorite book this year was the novella The River Has Roots by Amal El-Mohtar (who also wrote another favorite, This is How You Lose the Time War)

I also especially enjoyed The Cold Millions by Jess Walter, an novel about turn-of-the-century labor organizing across the pacific northwest and midwest.

I listened pretty incessantly to Ten Stories by mewithoutYou.

Ten Stories is an allegorical collection of songs that, at first listen, follows a winding narrative about a circus train crash in 19th century Montana.

The song Nine Stories feels most vivid in heavy Minnesota snow.

Late 1800s midwestern Americana has loomed large in my imagination for some reason this year.

$6 IKEA Kapplake lights make excellent desk backlighting

Thu, 01 Jan 2026 00:00:00 +0000

The IKEA Kapplake accent light works really well as a backlight.

Compared with the picture above, it looks a bit softer and warmer in person. The entire light is a toggle button, so I’ve put them toward the top and right side so they’re easy to reach.

(Yes, that is a UHK)

Behind the laptop stand

Behind the monitor

Box does not include a USB A power source, but surely you have a dozen of those by now.

Generate live albums from YouTube videos with ffmpeg

Fri, 19 Dec 2025 00:00:00 +0000

There are plenty of high-quality full length concerts available on youtube. It only recenlty occurred to me that I could just turn these into my own bootleg live albums.

Step 1: Manifest file

Make a new text file that follows this template.

artist: CHVRCHES
album: Live at Glastonbury 2016
date: 2016
source: https://www.youtube.com/watch?v=jIbRto7TQ30
---
02:45 Never Ending Circles
05:56 We Sink
09:54 Keep You on My Side
14:24 Make Them Gold
19:31 Gun
23:21 Bury It
26:59 High Enough to Carry You Over
31:32 Empty Threat
35:59 Recover
40:09 Clearest Blue

Step 2: Run the script

The script downloads the video with yt-dlp and transforms it using ffmpeg. ffmpeg even supports ID3 metadata tagging.

./annotate_v2.py my_manifest.txt

Results

Now you’ve got an album directory. You can drop a file called cover.jpg in if you want, but it’s optional.

➜ output git:(main) ✗ ls -T ./CHVRCHES/
./CHVRCHES
└── 'Live at Glastonbury 2016'
 ├── '01 - Never Ending Circles.mp3'
 ├── '02 - We Sink.mp3'
 ├── '03 - Keep You on My Side.mp3'
 ├── '04 - Make Them Gold.mp3'
 ├── '05 - Gun.mp3'
 ├── '06 - Bury It.mp3'
 ├── '07 - High Enough to Carry You Over.mp3'
 ├── '08 - Empty Threat.mp3'
 ├── '09 - Recover.mp3'
 ├── '10 - Clearest Blue.mp3'
 └── cover.png

Done. This was a one-shot with Claude Opus 4.5 and I’ve now made a dozen new live albums. I’m aware that you could embed the art direclty into the mp3 files, but I’m too lazy for that.

Navidrome setup

The “album” wont match anything on MusicBrainz, so IMO it’s best to keep them in their own siloed library away from Lidarr or any other metadata automation you might have running. Navidrome supports serving multiple root directories.

GitHub repository

https://github.com/subdavis/youtube-album-generator

rsync it to Navidrome

Alright, now push it to the sever.

rsync -avz --ignore-errors --no-o --no-g --exclude data/output/ user@host:/media/4tb/music_manual

Bonus: Audiotree

In case you didn’t know, you actually can buy audiotree sessions on bandcamp! My favorite so far is mewithoutYou from 2016 during the Pale Horse tour.

Using the Linkding bookmark manager as a CMS for Hugo blogrolls

Sat, 13 Dec 2025 00:00:00 +0000

I want the ability to easily share tagged links to this website, inspired by the blogrolls of yore (shoutout to sethmlarson).

This is a static hugo site, though. I have no interest in creating a commit just to make new links appear.

Linkding, my bookmark manager

Incidentally, I recently spun up a linkding instance to unify my bookmark collection and was delighted to discover that it has a public sharing feature. You can see my shared instance here.

This is the perfect tool for creating structured link data legible to Hugo!

There’s a convenient API and /api/bookmarks/shared does not require an auth token.

The same thing could probably be achieved with Linkwarden.

Adding a link

To share a link to this site, all I need to do is bookmark it with Linkding and check the “Share” option. The chrome extension is shown below.

Generating a Hugo page from Linkding data

Hugo generates pages from markdown and structured yaml data. I can curl my public bookmarks and write them out as json in the hugo data directory.

`./data/retweets.json`

curl https://links.subdavis.com/api/bookmarks/shared/ | jq > data/retweets.json

`content/retweets.md`

A static page definition

---
title: Retweets
type: retweets
---

Page footer content goes here.

`./layouts/retweets/single.html`

A page render template. Mine can be seen here.

{{ define "main" }}
<article>
 <h1>{{ .Title }}*</h1>
 {{ if .Site.Data.retweets }}
 <ul>
 {{ range .Site.Data.retweets.retweets }}
 <li>
 <a href="{{ .url }}">{{ .title }}</a>
 {{ if .tag_names }}
 <span>
 {{ range .tag_names }}
 <span>{{ . }}</span>
 {{ end }}
 </span>
 {{ end }}
 {{ if .description}}
 <span>{{ .description }}</span>
 {{ end }}
 </li>
 {{ end }}
 </ul>
 {{ else }}
 <p>No retweets yet.</p>
 {{ end }}
</article>
{{ end }}

Nightly GitHub Action

Action definition: https://github.com/subdavis/subdavis.github.io/blob/main/.github/workflows/fetch-retweets.yml
Run example: https://github.com/subdavis/subdavis.github.io/actions/workflows/fetch-retweets.yml

Runs nightly, and if a change is committed, a new build in CloudFlare Pages is triggered.

The result

I now have a “blogroll” style page thing at /retweets

The one-click Navidrome bandcamp purchase sync

Sat, 06 Dec 2025 00:00:00 +0000

Why is buying music so inconvenient?

It’s kind of insane that setting up a piracy stack is somehow easier than regularly purchasing and downloading music in a way that supports the artist.

Read this setup guide for example. Toward the bottom of that article, Nik documents his end-to-end purchase workflow. It’s horrible. It takes ~10 manual steps and 4 different websites to… buy an album.

There are powerful, convenient tools for torrenting. But if you want to buy things, the only platform not aggressively doing vendor ecosystem lock-in is Bandcamp. It has no consumer-facing API, its app kinda sucks, and there’s no good way to grab your purchases except one-at-a-time through the browser.

Navidrome and Lidarr

I do use Lidarr for some things, but I keep the lidarr library completely separate from my bandcamp purchases.

I ran into trouble trying to download bandcamp albums into a directory that lidarr controls becuase the file structure and naming lidarr picks are different. This breaks bandcampsync’s ability to determine whether or not it has already downloaded something to skip it.

Navidrome supports indexing multiple library folders, and will serve them all as if they were one librar. This works perfectly, so I keep them separate now. See the diagram above.

The one-click workflow

Until bandcamp build webhooks (fat chance), you need at least 1 click to trigger a re-scan of your purchases.

Buy an album on Bandcamp
Swap over to my Bandcamp Sync Flask instance and click “Run sync”. It pulls every purchase into /music_other
Navidrome discovers the new files. Lidarr is not involved.

Stuff should appear in your media player instantly. Bandcamp’s file downloads are already well-tagged so Lidarr would only get in the way.

The repo: Bandcamp Sync Flask

https://github.com/subdavis/bandcamp-sync-flask

Credit and kudos to BandcampSync which is the script this app wraps.

Comments

All this fuss about metadata seems like unneeded headache. A music library should be listened to on global shuffle, and if the RNG takes you straight from “Carolina in my Mind” to “Stupid Horse” then is that so bad? - Hastings Greer

How is our winter changing in the Twin Cities?

Tue, 02 Dec 2025 00:00:00 +0000

What does winter feel like

I sometimes hear about how winters are nothing like they used to be. Were they really worse in the 80s?

Memory is foggy. Averages and trendlines can be hard to relate to subjective personal experience. The average annual temperature in Minnesota has increased by 3.2°F between 1895 and 2024, but a ~0.5 degree per decade increase is not easy to feel year to year.

Still, the Twin Cities are experiencing climate change. That mostly means more variable weather patterns. Bigger swings, not just steady warming. Can we see that in the data?

Snowfall trends

Snowfall is probably the biggest factor for of how a winter is remembered.

There are some trends in first-order measures of snowfall, but they may not be what you expect. Both total snowfall and variability from winter to winter are increasing. Variability is actually down somewhat since an elevated period from about 1970 to 2010.

My rolling standard deviation is somewhat arbitrary, so I included both a 20 year and a 30 year to show sensitivity to the window size.

Something to keep in mind here is that regional seasonal snowfall is not normally distributed. The curve is skewed right, with a longer tail of very snowy winters.

My use of standard deviation above isn’t quite correct, so let’s try a box plot instead.

Let’s check how many snow days there were, and how much of the winter included ground cover.

There is a clear decreasing trend of days with ground cover, but it’s actually falling back toward levels from the early 20th century from a peak in the 50s.

I also looked at freeze-thaw cycles and “snow cycles”, the number of times snow cover appears and disappears completely, which I just made up. These have both remained relatively stable and very weakly correlated.

This is actually pretty interesting. Fewer days with snow on the ground, and fewer total days with snowfall, but the same number of melt cycles, implies that snow used to just… last longer? But with the same-ish number of snowstorms?

Ok yeah that seem right.

Temperature variability

Winter low temperatures are less stable and warming faster than other parts of the year.

In context, the global warming rate is about +0.36°F per decade since 1982. Northern Minnesota is warming even faster.

Lake Ice

Minnesota’s lake ice season decreased by up to 14 days over the last ~50 years. Compared to 50 years ago, the average July-August surface water temperatures on Minnesota lakes are 3.0 – 3.9 degrees warmer.

Source: Minnesota Pollution Control Agency (PCA)

Summary

Winter is changing in ways you might not expect.

Temperatures are rising at rates I would describe as both alarmingly fast and difficult to perceive.
The lake ice season is shrinking.
Snow patterns are highly variable since tracking began in 1880. A big bump in latter half the 20th century makes the last 30 years of decline seem steeper than it would be in absolute historic terms.

Footnotes and methods

You can do a lot using the DNR Minnesota Climate Explorer. For some of my own analysis, I verified my numbers against their tool.

Data and Analysis Tools

All the data and code used for this post is available on GitHub. I’ve mainly used Python with Pandas and Chart.js for the visualizations.

The Minnesota DNR provides histoirical daily climate data going back to the 1870s on their website. The downloads are all broken, but I was able to scrape them directly and explain how in the code repository README.md.

The 1970s pivot

Climate trend lines (like the ones above) intended to show deviation from historical patterns often begin in the 1970s without explaining why.

1950(ish) is considered the inflection point of human-caused greenhouse gas emmisions.
Until the 1970s, the effects of arisols in the atmosphere artifically suppressed those trends a bit.

The amount of CO2 in the atmosphere, measured continuously at Mauna Loa Observatory since 1958, is a startlingly smooth curve.

Disclaimer

I’m not a climate scientist or a statistician. I was interested in seeing second-order measures of variability that I’ve never really seen discussed in an emperical way.

Thinking about Generative AI

Sun, 23 Nov 2025 00:00:00 +0000

Long before we get to the point where a machine is a moral agent, we will have machines that are capable of suffering. - Ted Chiang

I want to write out some thoughts to revisit later. It can be easy to get revisionist about what you believed and when, so here is a list of things I thought were true in late 2025 after ~2 years of pretty extensive experience with frontier models, especially Anthropic’s Claude.

Some takes

AI is alarmingly good at all sorts of things. It is good at writing code, dealing with legal documents, talking through household repairs, and many other tasks.
We are not in a bubble comparable to previous bubbles, but the market is overheated. Many participants will suffer when their bad bets fail. A sector-wide collapse followed by another “AI winter” is not likely, but a recession that affects regular American families feels very possible.
Environmental concerns about AI and data centers can be pretty misleading. Environmental harms are important to mitigate, but environmental advocates seem overly focused on data centers compared to their relative impact or the marginal cost of reducing that impact.
As the US government actively sabotages renewable energy infrastructure projects, power demands from AI companies is both applying positive pressure toward renewable energy transition and keeping fossil fuel plants online longer than they otherwise would be.
Using AI to produce art, especially that which supplants human artists, is bad. Artists have a unique and important role in a healthy society. It’s clear that this is already happening, and I’m very worried about it.
Playing around and making art for personal use or amusement is mostly benign, though.
AI boosters are pretty universally untrustworthy. Most of the experts on this technology have conflicts of interest, and their incentives are not aligned with public good. We need more independent experts outside the AI industry.
AI skeptics often confuse material debates about the merits and capabilities of AI systems with larger philosophical or moral debates about capitalism. These are both important, but conflating them can result in motivated reasoning.
AI will, in the near term, accelerate wealth inequality. Skeptics are correct about this. Longer term effects depend on what we decide to do as a society.
AI will cause (mostly psychological) harm to humans at higher rates than other digital technologies. It will make harms from social media or smartphone use seem quaint. Society will mostly tolerate these harms and ignore them, just as it ignores the 1 million annual worldwide deaths from automobiles (40,000 in the US).
Large language models are not on the cusp of becoming self-aware or developing the capacity to suffer, but I am still worried that we might cause machines to experience suffering in the future.
We have avoided nuclear holocaust thus far by luck. Nuclear disarmament should be an urgent priority. All sides of the AI discourse should agree on this.

Disclaimers

Qualifiers like “maybe” or “likely” were omitted because I am uncertain about everything, and otherwise I would have placed them everywhere. My goal was to stake out some positions, not hedge against being wrong.

I am neither a skeptic nor a booster. I think these technologies are exceedingly interesting. The producers of LLMs are mostly untrustworthy, but the nuance has been lost. We usually don’t blame Facebook users when Meta does something shitty.

We exist in a dynamic period where challengers across domains could leverage AI against more lethargic incumbents. Civic and social good might result if good people would be willing to selectively wield AI in service of their desired outcomes. There’s plenty of reason to be angry at Meta, for example, but it doesn’t make sense to stubbornly refuse to use Llama, PyTorch, sam3, or React in pursuit of a better future.

Self-hosting Sustainably

Sun, 16 Nov 2025 00:00:00 +0000

At some point, every home server operator loses steam. After 6+ years running ~20 applications on a home server, I wish only for things to work properly and not waste my time.

If you’re anything like me, docker compose is probably all you need.

Here is my full configuration

My needs

Any conversation about the “right” setup should be grounded in goals. These are mine.

A bunch of network-reachable applications
Some reachable over public internet
Some behind SSO, IP Whitelist, or both
Running as <app>.subdavis.com
Plus DNS filtering via AdGuardHome

I no longer value arcane wisdom earned through suffering.

10 year old hardware

An old 2010s Dell Optiplex Core i3, 8GB RAM
Rehoused in a Fractal Design Node 304 Mini-ITX
With a 4TB media HDD
And a 2-Drive 2TB RAID-1 MDADM array for more important data
And another 1TB HDD for backups

Everything is computer docker compose

My basic philosophy is that I should be able to get this junk running on new or reimaged hardware in a couple hours if needed. I can apt dist-upgrade worry-free. Docker’s restart policy has never betrayed me the way systemd repeatedly has..

A little bit of systemd, though

systemd starts dockerd, which takes care of bringing up the compose stack on restart or container failure. Docker handles dependency constraints with grace.

One extra need is for the external disks to be mounted properly before docker even tries to start. fstab can create race conditions, but systemd allows you to set mount dependencies on services, which is quite handy.

[Unit]
Description=RAID1

[Mount]
What=/dev/md0
Where=/media/primary
Type=ext4
Options=defaults,auto,noatime,exec

[Install]
WantedBy=docker.service

Traefik is whatever, man

I use traefik. I learned it back when I had more energy for such things. I geniunely like its tag-based config because it keeps my docker-compose.yml modular and clean, but traefik config is verbose and annoying to troubleshoot. If I were starting over today I might try caddy.

Still, the Letsencrypt support and Cloudflare companion have never let me down. Often, my new hostname has registered and is ready to serve traffic before the new app has even booted.

Tired of fighting DNS

Maybe 6-8x per year, AdGuardHome goes down. Last year a slowly failing power supply lead to a lot of random system crashes. On this matter I have moved past depression to acceptance.

If you, too, feel traumatized by dhcp-option=6, I recommend giving up on your end user device discovery dreams and dealing with DNS at the router instead.

I run dnsmasq on my Edgerouter, redirect all DNS queries to my adguard instance wit NAT, and have a NAT rule enable/disable switch I can throw to take my AdGuardHome instance out of the critical path of internet access. I wrote about that here.

(The NAT rule is necessary regardless of your config because plenty of IoT devices hard-code their DNS anyway. Looking at you, Philips Hue Bridge)

Wireguard? 🙅 Tailscale? 👍

Wireguard config has consumed many hours of my life. The final straw was tweaking MTU while dealing with packet loss over mobile networks (T Mobile). I noped out and set up Tailscale.

No more innovation tokens

I started out using CoreOS which turned out to be foolish because it was abandoned about 2 years later.

I also used to do insane, truly unhinged shit with systemd. I believed that docker compose was not a valid substitue for an init system, whatever that means. Don’t be like me.

Unnecessary complexity is fine for learning. A particular choice may emerge out of a desire for novelty or fun or curiosity, but be clear-eyed, lest you confuse these indulgences with actual improvements. It’s perfectly reasonable to tire of tinkering with complicated setups. Using docker compose isn’t admitting defeat.

Keep it simple, stupid.

RSS - Really Simple Syndication

Sat, 08 Nov 2025 00:00:00 +0000

Podcasts are nice. They show up in chronological order, download in the background, and work the same in every app. In a world of algorithmic walled garden social “platforms”, this is a small miracle.

Why can’t more of the internet be like that? As Yair Rosenberg wrote in 2023, you can take back control of what you see on the internet.

It’s just RSS

The technology of a hopeful future actually came out in 1999. It’s just RSS. Every morning I fire up NetNewsWire. I pick something, open it up, and read it. Nobody tracks me. There are no cookie policies to accept. There’s no like button or comment section.

In the days before profit motives overwhelmed product design, humans understood that the media you wanted to access had very little to do with the “platform” it was distributed by.

Imagine Youtube videos and Instagram comments and new issues of your local newsblog all just appear in one chronological feed. The obvious problem with all this control is that nobody can sell ads or harvest your personal data if you’re safe in your own third party app (Outside of a web browser! Offline, even!).

Podcast freedom: fleeting

Even podcasts may not have this freedom forever as companies try to capture the remaining frontier. The New York Times is taking steps to force listeners into their own app. Spotify made major plays in the last few years to get exclusive podcasts behind their paywall. Podcasters now implore that you watch their episodes on YouTube, where they can more effectively monetize and dissect a larger stream of analytic data. Eyeballs, baby.

It’s perfectly reasonable to want subscribers to pay for your content. Substack and Patreon are positive role models for how this should work: You give them cash and they give you an individualized subscription link to import the feed in your own app. The ads you hear are at the discretion of the creator, not the platform.

Hunting for ghost feeds

Despite the incentives for companies to keep you on their site, looking at their ads and contributing to their engagement metrics, RSS doggedly persists. Idealistic developers have built RSS into most of the content management platforms (like Wordpress) that media use (mostly without paying). Some, like The Atlantic, even publish articles paywall-free, though the feeds are tough to find.

I’ve been using a chrome extension called Get RSS Feed URL that helps surface hidden feeds. It works in places you might not even expect, like Bluesky.

Conscious of content

Adopting an RSS-first means of consuming the web takes work, and the setup user experience can be rough. The exercise, though, can be enlightening. You might have a sense of how much time you spend on your phone in a day, but where is that time spent?

The content of my media diet have changed a bit in the last 6 months since I went all-in on RSS. It took real effort, but now, when I open my phone, I reach for my RSS reader rather than a browser or a social app. It feels like an improvement.

A few photos from Pittsburgh

Tue, 04 Nov 2025 00:00:00 +0000

Photos from RandyLand, Bicycle Heaven, and around town in Pittsburgh, PA. October 2025.

I want to get off Mr. Bones' Wild Ride

Sat, 01 Nov 2025 00:00:00 +0000

In June of ‘25 I shut off most of my social media accounts: Facebook, Instagram, and LinkedIn. I kept reddit under the happy delusion that I could learn to use it responsibly (lol) but I’ve made strides to curtail that as well.

Since then, I’ve often thought about what the experience of being “on the internet” is like, and how that’s changed since I began forming digital habits back in High School. My memory is relatively poor, so a habit-forming tool for capturing a bit of life is wonderful. Having my good intentions packaged up to generate ad revenue and give teenagers an anxiety disorder is less so. 10% of Meta’s revenue for 2024 came from fraud.

Still, I enjoy reading posts by people I’ll never meet. Putting stuff out into the world is fun. The Bear Blog discover feed is usually pretty wholesome, HackerNews is still rockin’, and my RSS app is chock full of honest work by decent humans. In the near future I’d even like to recommend some of them.

The prevailing form and style of most social platforms lack honesty and curiosity.

Brandon Davis

Amtrak and Skiing Whitefish, MT

Claude Code SDK is pretty wild

Beyond --print mode

Control handoff and tool use

A Toy Example

Takeaway

Renee Good was killed by ICE agents in Minneapolis yesterday

The reality

Other information

A final note

Thoughts for 2026

$6 IKEA Kapplake lights make excellent desk backlighting

Generate live albums from YouTube videos with ffmpeg

Step 1: Manifest file

Step 2: Run the script

Results

Navidrome setup

GitHub repository

rsync it to Navidrome

Bonus: Audiotree

Using the Linkding bookmark manager as a CMS for Hugo blogrolls

Linkding, my bookmark manager

Adding a link

Generating a Hugo page from Linkding data

./data/retweets.json

content/retweets.md

./layouts/retweets/single.html

Nightly GitHub Action

The result

The one-click Navidrome bandcamp purchase sync

Why is buying music so inconvenient?

Navidrome and Lidarr

The one-click workflow

The repo: Bandcamp Sync Flask

Comments

How is our winter changing in the Twin Cities?

What does winter feel like

Snowfall trends

Temperature variability

Lake Ice

Summary

Footnotes and methods

Data and Analysis Tools

The 1970s pivot

Disclaimer

Thinking about Generative AI

Some takes

Disclaimers

Self-hosting Sustainably

My needs

10 year old hardware

Everything is computer docker compose

A little bit of systemd, though

Traefik is whatever, man

Tired of fighting DNS

Wireguard? 🙅 Tailscale? 👍

No more innovation tokens

RSS - Really Simple Syndication

It’s just RSS

Podcast freedom: fleeting

Hunting for ghost feeds

Conscious of content

A few photos from Pittsburgh

I want to get off Mr. Bones' Wild Ride

Beyond `--print` mode

`./data/retweets.json`

`content/retweets.md`

`./layouts/retweets/single.html`