Sam πŸ’›

Sam πŸ’›

@samhh.com

Human of a few decades. Loves animals and functional programming and pastel colours. https://samhh.com https://github.com/samhh Working on @gitbutler.com 🩡

216 followers426 following405 posts

It's been a few months since I vibecoded some personal software, back then with Sol 5.x. I'm doing some more this weekend with Opus 5.5 and it's surprisingly more capable, and isn't eating up my usage (Max 5x) like I expected/remembered.

I'm using the projects feature so I mostly just talk to the one coordinator with several things on the go. This is really nice if you don't care about the quality of the underlying code.

It's set up to push apps straight to my phone (and desktop), and can push to TestFlight for my partner as well. We're starting to use hosted services for some stuff like push notifications. Ideally it would have its own isolated accounts. πŸ€”

Sam πŸ’› reposted

Pumpkin spice lattes have returned to my local woodland cafe. And with that everything feels right with the world.

It can be interesting to read how the AI labs are doing agentic development. It’s even more interesting when they document the same limitations the rest of us experience, in which case I question how it’s viable for them. If you don’t have an unlimited budget and a golden goose, is it still viable?

Finally catching up on the Mercurial talk I missed at Git Merge. The data model seems very elegant and intuitive. Topics sound a little like how I have envisioned reviewing and merging (sub)sets of changes, a bit closer to Gerrit.

Linear’s mobile app emulates Liquid Glass in the bottom navigation rather than using the real thing. It’s icky.

iOS 27 and friends are about to drop. I’ve been running the betas since the early developer releases. This is the smoothest and most reliable the operating systems have felt since I adopted the ecosystem five years ago. I wish every release would balance features and polish like this.

Astra has generated some truly trash code for me this week. It does feel as though they’re optimising for replacing SWEs in order to facilitate cheap technically-functional junk, rather than enhancing SWEs to produce better output faster. Better lock-in that way too.

Not sure I can remember Sol/previous being as gracious as Astra here after multiple rounds of disagreement. (Yes I know I'm anthropomorphising too much.)

It's real nice when debugging libraries to have your agent spin up an ephemeral single-file HTML repro that even renders in-app.

I don’t really live in the terminal any more, but anyone currently feeling disillusioned with Neovim should try Helix. Break free from the Vi family!

It took me nearly a decade but I’ve finally learned how to forage and start a fire in DayZ. I am however now fighting off illness from eating baked wolf steaks. The game is oddly serene.

The Codex Remote -> push to iPhone over WiFi setup I have going is very effective for simple apps. I can do it all from my phone. I don’t know why we have Image Playground and not something to generate apps instead, in the spirit of Glaze. I suppose generating WebExtensions is the first step.

Codex can spin up a working iOS dev loop (remote, even!). Codex cannot get simple links to work within a Swift app. Flashbacks to attempting macOS dev a few months ago. Maybe the path of least resistance is React Native.

How did I only learn today that the mini floating keyboard in iPadOS supports swipe typing, like on iOS 😍

UK govt: if you send emergency alerts to my phone about things that are not emergencies and do not concern me (wildfires far away from my city) I will simply turn off those alerts. Cmon

To be clear, this is as relevant to me as tornadoes in Texas. This system only works if it’s reserved for true emergencies with appropriate geolocational scoping.

After a year of placing my iPhone Air face down unprotected, it has no visible scratches. Dropped it plenty of times too. Ceramic Shield 2 is more than marketing fluff!

I'm experimenting with treating the harness as my primary development app, dropping down to my editor only where edits are actually necessary. That's multiple times per day, but a minority of my time in the day.

I currently have three different views of code open at the same time, not including any forge tabs in my browser. I doubt that'll be the case in a year or two.

To my surprise, being Codex app-first, I'm actually messaging multiple threads at once pretty often. You can kick off agent investigation on something you've just noticed in the background without having to completely put it off.

I've switched from Handy to native macOS dictation. I don't know if it's improved in macOS 27 or I just didn't give it a fair shake before, but if all you need it for is transcribing to an agent then it's absolutely sufficient; LLMs are extremely good at parsing meaning from typos.

I've long resisted allowing agents to touch my version control, but I'm relaxing on that. GitButler's CLI skill makes agents very helpful at resolving conflicts, splitting commits, etc. Principally the oplog is always there (also in Jujutsu's case) if the agent does something stupid.

Agents with access to your forge can also resolve e.g. flaky CI E2E tests end-to-end (ha). There's typically little value in hand-writing the commit message and PR title in these cases. In the meantime I can work on something more impactful.

I'm finding myself writing bigger, more traditionally-atomic commits again. This actually isn't agent-specific, though it's relevant there: we need better review tools. Reading a long diff like a scroll feels antiquated.

I've dropped Codex 5.6 Sol down to medium from xhigh and am having a much better time. LLMs are still pretty stupid so the speed is more valuable >95% of the time. I haven't tried Fable yet, mind.

Reading/scanning code is more important than ever. Do any editors display dynamic metadata above signatures? e.g. a recursion label, approx big O, purity, etc without having to read the function body. Maybe in the future local LLMs will be fast enough for this.

I'm up to >1B tokens per month with Codex. Though that's using Sol xhigh for everything in a vain attempt to rein in its stupidity. I do wish the feedback loop was faster.

Sam πŸ’› reposted

I'm looking for a new role as a lead frontend/TypeScript engineer. πŸ‘‹ 13 years of experience and I formerly led the web team at Unsplash. My CV: drive.google.com/file/d/1Hp7w... If you know of any roles, please reach out. I'd really appreciate if you could share. Thank you!

I quite like my Eight Sleep so far but I can’t get over such an expensive device shipping without WPA3 support in 2026. I’m doomed to forever have two SSIDs.

From a recent Fly blog post, advocating AI everything: β€œI want a world where my friends and family can make computers do exactly what they want, without waiting for me to build everything for them.” A cute sentiment that falls apart on closer inspection. Sometimes this industry is so unserious.

Using agents to crank out features of comparable quality in a fraction of the time feels empowering. Likewise being able to dig into a new codebase and ask human questions and usually get good answers. As long as you pick up on when to push back against the plausibly wrong output!

Noticing in Zed with agents I frequently don't want "queue and send", but rather "queue for later". It can take many messages to resolve something which blocks sending the message for the next thing I want that I've already written. At the moment my workaround is the clipboard.

Coding agents are great, but they seem more impactful than they are because they've replaced the majority of typing which occurs via a) more or less copy/pasting documentation and b) following compiler errors. When you are in more advanced weeds their value falls off a cliff.