A thought on LLMs as a co-designer —
In the old world, a frustrating thing was that I could see exactly the layout I wanted in my head, but could never make it perfect on the computer screen because there'd always be some CSS bullsh*t where I'd spend 4 hours trying to get like 5 pixels straightened out, and even if I got it fixed, it'd happen 20 more times.
I was in the business of blogging and not building blog designs, so I'd go for the 80-90% mark, declare "good enough", and return to my main quest line.
Now, Codex/Claude is like a best buddy who doesn't mind spending time brute forcing this abominable language (no designers, CSS is not good: it is bad, always was bad, and will continue to be bad) to get whatever needs to be done, done. With their help, I can get the design exactly as I envision it down to pixel perfect precision.
The only `.github/pull_request_template.md` you need. Less (an empty file) is also good. More is too much.
I wish I'd saved a screenshot of the templates at Stripe and Snowflake. Dostoevsky-length volumes full of context no one read and nanny-ism that was universally ignored.
This is a place where the precautionary principle bites —
It's really, really easy to argue that an exhaustive template full of hard-won wisdom and encyclopedic catalogization of all known caveats and warnings will help improve the safety of future deploys.
But then later, once it's crystal clear it's not working because even if someone read it one time, it's now just background noise that adds friction to every pull request from now until the end of time, it's really hard to argue that it should be torn out.
What, you don't like SAFETY? What's wrong with you?
This account is genius — so on the line that 80% of the population can't tell it's satire (or is it???). Exactly where comedy shines brightest.
xitter.cf/compliantvc/status/211…
This is why Europe is the best place to live
Our financials are incredibly stable
EU citizens can store their money in our stock market and trust that it's completely safe
Look at this: No one lost money over the last 18 years
Stable currency. Stable markets.
Meanwhile Americans lose money to inflation every month.
The European dream is alive and well
Nice article on the suppressive effect of subtransactions on a job queue. I'm regularly confused about `MultiXact` and why it has such a negative effect, so it was helpful.
And BTW, if you're having trouble with this, try River! Subtx-free since 2026.
incident.io/blog/how-savepoi…
A lesson many of us could stand to learn: perfection is not a prerequisite of success.
An at-risk job of the LLM age is the do-nothing hall monitor who criticizes everyone else's work because it falls short of some ivory tower ideal. You know the guy.
xitter.cf/brandur/status/2111170…
I'm contracting at a company this week about their API and very modern Rust stack. I asked, "So I assume the OpenAPI spec is generated from API DSL/structure?"
Their engineer: "Nope! Periodically we have the LLM go over our code changes and make additions to the spec."
Not long ago this would have greatly offended my sensibilities around code hygiene and correctness. It still does a little, but the more I ponder, the more I think: it's probably fine?
The spec could get behind, but you keep a high order conflict-heavy generated artifact out of the firing line, and in practice the former probably doesn't matter.
I'm contracting at a company this week about their API and very modern Rust stack. I asked, "So I assume the OpenAPI spec is generated from API DSL/structure?"
Their engineer: "Nope! Periodically we have the LLM go over our code changes and make additions to the spec."
Not long ago this would have greatly offended my sensibilities around code hygiene and correctness. It still does a little, but the more I ponder, the more I think: it's probably fine?
The spec could get behind, but you keep a high order conflict-heavy generated artifact out of the firing line, and in practice the former probably doesn't matter.
River has a new home page!
@blakegentry wanted an homage to the old school Heroku login page circa 2011.
Fun detail — the landscape is day or night based on darkmode.
Inclusion of sakura tree boosts E2E work perf by 10%. Or maybe I made that up.
riverqueue.com/
Google Docs' new markdown support is awful:
• Markdown is just a shortcut to convert plain text to formatted text.
• Once formatted, text can't go back to markdown.
• Copy/paste copies formatted text.
20+ years coming and a catastrophic flop. Very, very disappointing.
Been thinking about this meme multiple times a day for weeks.
I'm sent something daily that would've put jaws on floors in any previous year, but now just blends into the background. The challenge of the coming age won't be to create product, it'll be to get anyone else to care.
Huge River ship (biggest ever?): River, the polyglot queue.
We've transliterated the project to Rust, TS, and Ruby and refined implementation and APIs endlessly. The result is the full River core, rewritten from the ground up. No Go dependency required.
riverqueue.com/blog/polyglot…
Packages in each ecosystem:
• Rust: crates.io/crates/riverqueue
• TypeScript: npmjs.com/package/riverqueue
• Ruby: rubygems.org/gems/riverqueue
A reasonably easy Dependabot fix would be to allow security updates to group with everything else.
Security teams would hate this because they're religious zealots, but just because GitHub sends a security fix immediately doesn't guarantee a new release immediately. Far from it.
The more I use Dependabot, the more I want to go back to a custom dependency update job that operates on a strict schedule, and only once/period.
The insane amount of noise/churn Dependabot creates makes the service only borderline worth it at best.
brandur.org/fragments/mostly…










