Here's a certified human written post from the one and only Yvon Chouinard about tumplines. Patagonia actually used to sell a tumpline compatible with some of their packs briefly in the 2010s!
I've been a trackpad proponent for years, but with Steermouse I've been able to dial in my MX Master so perfectly that I find myself missing it when I travel.
What really took it to the next level for me were adding chord mappings. For example I can hold right mouse button + scroll the thumb wheel to adjust volume. So convenient!
Your point on the harness is interesting. How do you distinguish characteristics of the model from characteristics of the harness?
In the early days I feel it was more apparent. You would frequently see the model making failed tool calls etc.. but now that feels so rare. I'm not confident I can perceive whatever shortcomings of the harness remain.
I'm not talking about model performance. I just mean the UX of Claude Code. I'm trying to use pi but there are so many paper cuts. Of course you can configure everything but that's a ton of work. Claude Code has pretty good defaults.
Bit of a tangent but at work we have GitHub Copilot and the VSCode harness is somehow night and day better than whatever happens in the IntelliJ plugin. Aside from having better features, for some reason prompts seem to be cheaper as well.
I was using Context back when it was still Brand.dev. I found it to be a great product- one of those rare APIs that immediately made a problem I had disappear. Had it in production within an hour of signing up
Agents need clean/current context from the web, and this is the best way I’ve found to give it to them. The internet is clearly moving in this direction: companies are starting to realize their sites need to be legible to agents. Some are already adapting but many haven’t yet. Context feels like an important part of that transition
Yahia is a great builder. His pace of expansion has been impressive, excited to see where he takes Context.
Exactly right. Climbers care a lot about the ethics of an ascent. It’s interesting how much those ethics have changed over the history of the sport.
One of the core ideas is that later climbers should respect or improve upon the style of the first ascensionist. e.g. if a climb was first done using siege tactics, then doing it in a single day is celebrated. But making a climb easier or safer after the fact is much more controversial, because it can feel like changing the nature of the route itself.
Snake Dike is a good example that’s flared up recently in the climbing world. It’s a classic, relatively easy route up Half Dome, and many climbers free solo it. But because it’s a face climb, protection mostly comes from bolts drilled into the rock. The first ascensionist placed very few bolts, which left long runouts and real consequences if you fall.
To many old school climbers, adding bolts to Snake Dike is disrespectful because the risk is part of the route’s character. Their view is basically: don’t bring the mountain down to your level. The new generation of climbers don’t seem to feel that way at all - they think you shouldn’t have to take unnecessary risk to climb a classic route.
I agree with the old school climber's philosophy. It's almost like the old school emphasizes the sense of adventure, and the new school emphasizes the sense of experience via touring (which in my mind are two different things).
I like this - I think you're not too far off of what's popular these days though. I think similar functionality can be achieved by using the "hook" functionality in claude code / codex.
I've been using both on a Rust codebase and have found both work fairly well. Claude code is definitely more capable than Gemini. What difficulties have you had?
The biggest pain point I've had is that both tools will try to guess the API of a crate instead of referencing the documentation. I've tried adding an MCP for this but have had mixed results.
It might be that we have multiple features in our codebase and Gemini seems to struggle understanding that it needs to be aware of #[cfg(feature = "x")] and also that if it's trying to run things, it might need to specify the feature.
And yes, when they guess APIs, it's highly annoying.
Congrats on the launch! This is really cool - one of the applications of LLM I find most compelling. I've seen so many back office processes that have hundreds of steps, are incredibly error prone, and traditionally couldn't be automated due to API limitations. Solutions like Skyvern are going to supercharge businesses that have had historically low margins due to the number of humans required. (Not as a replacement for a human, but as a force multiplier)
The most fascinating part is how tough that work really is. Everyone we've talked to loathes the manual stuff, but until a better solution comes out, you have to allocate X% of your time to it
Only if you require text message based two factor. Password managers like 1Password allow you to store your OTP within them and share that + the password internally within your team
I've set up multiple times a phone-to-Slack proxy for this exact reason. In my case it was a VoIP number, but if that's blocked, Android has many SMS-to-webhook apps and even entry-level industrial LTE routers generally have this feature so you can use a real SIM card.
https://www.patagonia.com/stories/culture/design/on-tumpline...
reply