Pi agent's /tree feature (see also /fork) is neat, basically take your linear conversation and make it a... tree!
Example: ask for a thorough review, fix n at a time, /tree back to the review and go through another set. Could do the same for implementing a complex feature with summary on return.
Shortcut app steps replicating a quicklink to open Nix package search
You can recreate Raycast quicklinks in macOS 26 Shortcuts like so.
Spotlight kind of hangs onto the output sometimes for some reason, even if you explicitly "stop and do nothing" at the end. But hey, like this you get syncing for free 😉
Anthropic restricting how you use their subscription (re: harnesses/agents), and seemingly reduced usage limits, is pushing me to explore alternatives. Let’s see if there’s really a moat.
Realising as I type this that it’s similar to the problem businesses are having where AI is just helping them move in the wrong direction faster.
It’s that associated lack of thought that’s the problem: have you really thought through all the edge cases and implications? Have you taken a step back to really think about the problem?
Noticing a trend of so many cool new tools popping up that you realise have only been active for a few weeks and were almost certainly vibecoded with minimal thought.
cc @mlegenhausen.bsky.social I'd suggest a local one like this instead of what I was using when we talked about this a few weeks ago.
The Pi agent is interesting, it's a shame Anthropic have locked down Code subscriptions for other harnesses.
Particularly intrigued by extensions like this one that adds LSP integration: github.com/apmantza/pi-...
I then have macros for spawning Claude, Handy, etc on my Keychron keyboard's M1-5 keys. Much nicer than memorising a bunch of combos.
There are a bunch of macOS apps for this, I suggest using Handy and following its guidance on which model to use. I use Parakeet v2, which I'm finding both fast and accurate.
Starting to lean into using dictation when interacting with LLMs. You can just talk and let the output be imperfect, it'll be fine.
Just migrated my PDS to @eurosky.social, that was easy! Feels nicer having my data stored in Europe, even if it's all public anyway.
Managing my dev configs in Nix has become rather intractable with the speed of change around AI tooling. I want the shiny new things!
Two broken pairs of AirPods Max, one Lightning and one USB-C. Just noticed they both lasted almost exactly 1.5 years. Think condensation from open windows is killing them. 😫
Skill issue, you just need to generate larger chunks of incomprehensible code 😎
git bisect with some clever maths for when something is flaky or nondeterministic, very cool github.com/hauntsaninja...
I’d suggest doing jj new back to a known good point and organically branching off there, then abandoning unwanted experiments at the end.
It does the same with React when usually that’s the wrong approach. It’s trained on everything which in aggregate isn’t very good code regardless of stack.
Also, every single publication talking about Playlist Playground failing to observe it’s currently US-only.
iOS 26.4’s Apple Music changes amount to it largely ignoring light/dark theming, so probably don’t open Favourite Songs after dark. 😬
Today's reminder that LLMs can't reason:
> You have five active Thread Border Routers across two ecosystems (Apple + Google + Aqara).
I mean it’s taken a dozen reprompts to mostly fix the cursor issue in SwiftUI, and it’d be challenging to test for.
Continuing my vibecoding experiment, this time with Claude Code.
It can spin up whole new features with ease, but it'll repeatedly fail to render a pointer cursor on a button.
The key is giving it tasks that it can write tests for. You wouldn't however expect to need to test cursor states...
Same here. Quad9/UK if it helps.
Let’s not pretend like conflict markers aren’t frequently incomprehensible.
I’d tend to agree, though I fear the gap is where testing and arbitrary choice can make that looser specification technically functional.
It’ll be slow and unpleasant and inaccessible but it’ll potentially work and be far cheaper to produce, which our economic model will incentivise.
New blog post: "A sufficiently detailed spec is code"
I wrote this because I was tired of people claiming that the future of agentic coding is thoughtful specification work. As I show in the post, the reality devolves into slop pseudocode
haskellforall.com/2026/03/a-su...
Right, yes, this. Zed doesn't currently expose /effort for Claude Code so I think I was bang on with this one. I'd guess both in quality and speed Codex's "medium" is roughly Claude's "high".
Finally have access to Claude Code after using Codex for a while.
So far it (Opus 4.6 1M) seems a bit dumber ("low effort"?), but substantially faster to respond and do stuff. On balance, since I use it primarily for fuzzy text manipulation and as a second pair of eyes, I'm enjoying this change.
I appreciate how much further this idea can go with robust and verbose testing and specs, but I pity a future in which we have to live with even worse software than today's norm because it'll be cheaper to produce and quality tends to lose in this race to the bottom.