Posts for AI

Monday, 20 Apr 2026 10:44 pm

Claude Code for fell running

I’ve got a big run coming up, and I want to be prepared, so I created a GPX file to load onto my watch so that I can follow a little red line across the Yorkshire Dales.
A little red line alone is a big help, but a bigger help would be to have the checkpoints on the watch as well, so that when it’s been dark for hours and I’m stumbling through peat hags, I can know that I’m only 2 km away from Capplestone Gate. Or whatever.
The problem is that manually adding checkpoints to a GPX file is a tedious prospect in Garmin Connect. The solution is that I have a monthly subscription to the Do Tedious Things For Me robot.
Claude Code made short work of the GPX file: I pasted in a screenshot from the event handbook listing the OS grid references for each of the checkpoints, and it parsed those checkpoints first into a CSV and then interpolated them into the GPX file itself. Names and lng/lats, and, with a little further prompting, elevations for each of the waypoints.
I uploaded this file to Garmin Connect but found that some of the waypoints were missing; Claude discovered that the grid references provided (to 100m precision) weren’t always exactly on the plotted route. It tweaked the file; I reuploaded: success.

Claude AI Running The Fellsman
Friday, 27 Mar 2026 6:33 pm

Thoughts on slowing the fuck down - Mario Zechner

And just as if I'd called into the mist with my previous post — Mario Zechner's thinking returns my own voice to me:
We have basically given up all discipline and agency for a sort of addiction, where your highest goal is to produce the largest amount of code in the shortest amount of time. Consequences be damned.
Mario's solution is to use agents for limited-scope or non-"mission-critical" work, and to handwrite everything else. I think the handwriting horse has escaped the barn (or, uh, whatever the metaphor is; it's been a long week) but keeping a closer eye on your agents is certainly the right move.
I keep thinking of that "a computer cannot be held accountable" directive and it seems every day to be a better light by which to find one's way through the IDEs of the future. I refuse to commit code, agent-written or otherwise, for which I cannot totally account: this helps keep my mental models accurate so I can quickly fix things when they, inevitably, go wrong.
"Thoughts on slowing the fuck down", Mario Zechner

AI Mario Zechner Code
Friday, 27 Mar 2026 11:28 am

The code agent orchestra - Addi Osmani

I usually think that Addi is pretty clever, and it’s clear that he really knows his stuff — and it’s cool that all of these tools and methodologies exist, but I can’t help but think that this is not useful at all.
The parent orchestrator only talks to two agents, keeping its context clean. Feature Lead A gets a brief like "Build the search feature" and decomposes it into Data, Logic, and API subagents on its own. The parent never sees those details.
This doesn’t solve any of my problems, which all pertain to expertise, complexity management, and deep knowledge of systems. Whereas a lot of these breathless AI accelerationist thoughtpieces basically assume that everyone is just trying to build the same greenfield project management applications, over and over, all of the time.
"Build a reusable card component." "Integrate with this API." "Build the search feature." What if the search feature was already built, 8 years ago, by someone who has dropped off the face of the earth, and for some reason stopped working earlier this week?
I’d like to think that I’m reasonably bullish about using AI for programming — I’ve had some success with using Claude to increase the pace of development at work — but this just looks like it generates heaps of new code that no one knows how to understand, not even the agents themselves.
"The Code Agent Orchestra", Addi Osmani

Addi Osmani AI Claude
Sunday, 15 Feb 2026 7:33 pm

Cognitive debt

From Simon Willison, in a post about Margaret-Anne Storey’s concept of cognitive debt:
I no longer have a firm mental model of what they can do and how they work, which means each additional feature becomes harder to reason about, eventually leading me to lose the ability to make confident decisions about where to go next.
This is a similar problem to one that I’m dealing with in a large legacy system at work: the people who designed the system are no longer around, and many of the decisions made in the interim have been bottom-up solutions to transient problems, rather than cohesive approaches working with the grain of the application.
I suspect that the advent of LLMs for programming is not going to resolve the issue.

Code AI Work
Saturday, 31 Jan 2026 8:44 pm

AI lowers the effort barrier

Maybe this is obvious to everyone but me, but the biggest unlock to AI seems to be less “AI helps me do things that were impossible before” and more “AI helps me do things that I just didn’t have the time/effort/attention/care to do before”.
See e.g. this post from Simon Willison about a theatre lighting cue software developer:
This latest TikTok describes his Claude Opus moment, after he used Claude Code to build a custom lighting design application for a very niche project and put together a useful application in just a few days that he would never have been able to spare the time for otherwise.
Or this post about a clip from Andy Masley referencing a podcast with Emily Bender and Robert Wright and Alex Hanna:
Wright: What I find valuable is, if there’s something I just know nothing about, I mean like, theories about why Rome fell or something, I find it a much for efficient way to get a sense for the landscape […]
Bender: […] Oh, go to the library, talk to a librarian!
Wright: Yeah but it’s so much slower! It’s so much slower!
Or my own experience diving into an open source Go codebase that I’m less familiar with to fix a bug that I was seeing — I wouldn’t have done something like that previously because just getting my bearings would have taken hours.
Or trying to address the shit with which the zone has been flooded: Donald Trump claiming at Davos that renewable energy in the UK hasn’t brought energy prices down — this claim looks true! Energy prices are higher than ever (and it affects me personally). Normally I'd dismiss this out of hand; I know I don't have a clear picture of the situation and he's using a shallow truth to tell a deep lie. But instead I loaded up Claude and told it to go off and produce a bunch of research about why energy prices are high despite the rollout of renewable infrastructure across the country. It's not the whole picture, but now I can mostly debunk the shit in realtime — and I would never have been able to do to that before.
All of the above was possible before AI, and remains possible without AI — but the effort required has drastically been reduced.

AI Andy Masley Simon Willison Go Environment