> ...providing context files does not generally...
# self-promotion
t
...providing context files does not generally improve task success rates, while increasing inference cost by over 20% on average.
https://autonomousapps.com/blog/agents-md-expensive/post/
l
The problem is that most folks think "more is better" in those files...when, in fact, the opposite is true. I like my CLAUDE.md files, but i do find claude just ends up ignoring them when they get too long.
t
people absolutely think "more is better" for those things and I cannot understand why. It's contrary to every practice I learned in software before now
l
I think it's easy to understand why from my perspective because to me, software engineers have always had that mindset. Unlike hardware engineers who work against real, fixed constraints like die area, power/thermal budgets, register widths, specific cache sizes...When you're constrained by physics and manufacturing, you learn discipline early because there's no fallback. Software never had that forcing function. Moore's Law bailed us out for decades so if your code was slow or bloated, just upgrade your hardware and fix it. That's how we ended up with garbage collection and all that. Cloud autoscaling made us even worse....and now...AI...don't get me started!!!!! If anything AI has a perverse incentive to use more tokens and make your pay for more context with giant CLAUDE.md files!!
make you pay*
LOL it's big software.
t
I can't argue with you! I've personally always tried to be as minimal as possible. Maybe because I do a lot of build work with Gradle and local machines are always running out of memory! And for a while anyway, Android devices had an "embedded" quality. I think they still do, especially if you have customers outside of the US or EU with lower-end devices. I used to have to care a lot about whether my apps would work for someone with a cheap Android device.
l
👆Respect.