What it feels like to code in 2026
I am horse trainer with a magic wand, and the only limit for how fast I can make my horse go is my own imagination.
So...
I've designed bespoke electrical compression horseshoes that both propel the horse forward and recharge when the horse slows down
I have my horse on horse-sized doses of all the peptides Brian Johnson takes, plus a few experimental ones
I'm prototyping an F1-style back wing to generate downforce. The computer simulations show there's gains to be had, but Claude politely refuses when I ask how to attach it to the horse's ass.
And oh by the way if the above things sound crazy? Check out #fastHorseIdeas on x.com where people are doing things even I (a pretty out-there horse trainer) find offensive and shocking...but also might add to my workflow if enough people keep talking about them
User Assumptions for a world with no LLMs (or 2024-era LLMs)
One genuine surprise of becoming a professional software developer was learning just how dumb we assume that average user is.
Never show an error message beyond "Go talk to your administrator/support!" Tool tips. A pop-up modal telling the user how to use tool tips. And make that button bigger! Make it flash, too. Not so much that it could trigger a seizure...whatever that threshold is let's aim for riiiight under it.
User Assumptions in 2027
How close are we to a world where instead of assuming our user is a country bumpkin chewing on a long piece of hay staring at an old beige CRT monitor, and start assuming that our user has:
a frontier LLM co-pilot with access to our app's excellent documentation + OS-level stuff so it can take screenshots & read error reports without any help from the user
a PDS (or equivalent) with basic table-stakes data we can request access to read/write
something like wispr-flow so they can quickly & informally throw unstructured data back at our app
We can still picture them as the country bumpkin. They still won't know what a cache is. But with those three things, what kinds of software could we make? Would it even look anything like software?
A wild guess for Software 2027: What's halfway between a protocol and an app?
Notice the banner image for this blog post.
I did not generate the image of the horse for you. Why? Because your ten fingers work just as well as mine. Do you want to see that horse? Go nuts!
Once I generate that (cursed) image, that's it. That's the horse. That's the only horse there can be. But if I gently point out the possibility of the horse...the ball is in your court.
I keep think about an interaction I had at a meet up where a guy showed me something he vibe coded. A lot of bad LLM integrations (including thing I've made!) fall into a category where they limit what you can do without giving you enough benefit to compensate for that loss of control.
Sometimes you absolutely want to trade some control for convenience/guidance. But there should always be a reason.
Good Software in 2027 is going to have to pass the test of, why can't I just have Claude do this instead?
And so I wonder...if what we (humans) build will be more like the ffmpegs/ATProtos. Kinda like libraries, but not optimized to aid human coders. Optimized instead for pointing workflows towards specific classes of solutions. We still make the lego blocks. Claude builds the tower.
Just a wild ass guess.