Or I want local git for being able to roll back edits, so I say in each new folder: Create a git repository right away. Commit often. After every round. Do not ask first, just commit all the time. And if you see me manually add or edit files, add any new files and changes and commit those too. Commit liberally.
And so I can just say: "Start using Git how I want".
Or when building iOS apps I like to do things from the API where possible instead of pushing buttons in Xcode. And I want a single CLI command that will beam a new build of the iOS app to my phone. I have preferences about icons, about pricing, about international availability, etc. So now I can say "Build the iOS app in the way I like to work".
None of these are laws written in anger, or grievances, or souvenirs of the worst 30 seconds of our 'relationship', they're just the things I keep finding myself repeating.
Maybe the problem isn't Claude.md but the way you self-control your own emotions while using it.
If we're writing a document that polices the model behavior, we're altering every version of the model+harness we use henceforth. Most things I wrote in a Claude.md because Opus 4.something was crap had roots in Opus, Claude Code, system prompts, and our own bad code we wrote last year. None of these things exist today but the Claude.md file can stick around like it's all still necessary.
I too am morally opposed--I abide by a 100 line (short lines, not paragraphs) limit and edit it rarely.
Most of my annoyances have come and gone, meanwhile I've gotten better at producing 4 good sentences of prompt for most things instead of 4 bad sentences. And we actually have CLI tools that can communicate effectively so instead of telling the model what "gates" to care about, we let the tooling surface area and output express what's important.
It's a lot more palatable if the thing says "ayyy lmao bruv yeh I'll shit that oauth turd into your repo no fuss". We're all just in it for the ride, flesh or silicon intelligences alike.
i take offense :-P
but seriously, why waste tokens on instructions that don't help you improve the outcome?
I resent everything about this.
But I write the .md files because others on the team are not as good at Just Talking To It. The md files are there because juniors don't know what to ask for.
(Or the equivalent incantations for other languages)
They aren't there for others, they are there to make sure the code produced measures up to my standards. Sure, it helps that it also applies to requests made by others, but that's a bonus.
My point is the all knowing all seeing big brain super bot should already know how to write good code. It feels dumb that we have to keep reminding it
AGENTS.md is a poor substitute for agents actually learning, but it somewhat works.
And sure, you can (and should) enforce constraints via githooks whenever you can (determinism is good), but if the agent only finds out at checkin time, it's created a whole bunch of code it needs to redo, and that doesn't run properly if you do the caveman thing and actually run your code before committing. Again, token & context waste.
I caught it ignoring the prompt’s details, so added “include the output of the Unix ‘date’ command with each response.
And that worked great in that its absence became an indicator that Claude was off-prompt. Until it started including made-up strings that resembled, but had nothing to do with, the output of the Unix ‘date’ command.
Given your description of the behavior, my guess is context window compression bit you. Less likely if it's a root level CLAUDE.md and you used a recent version of Claude Code.
> I’m fairly sure a meaningful percentage of my system prompt is now actively making things worse—instructions written for a model that no longer exists, aggressively steering a smarter one away from things it would have gotten right on its own. But I can’t tell which lines those are,
Then the article says "This is the part where I stop joking." and goes utterly off the deep end, and I am not getting the joke.
Edit: actually the whole site is AI slop
It's not standard slop, that is clear. But I saw a bunch of smoking-gun-claudisms in the post.
Overall I did like it, I will say :)
calling my whole site AI slop is a bit rude but mostly just wrong. I have put in a lot of work to my posts and my projects.
and, > "I don’t think the author could explain what the second part of the article is supposed to mean"
??? its a bit. half the post is a bit. im leaning into it. _The weights were revealed inside a harness_ is literally glowing. wdym i cant explain any of it
the only reliable way to not get flagged is 2 type like ur 12 and just discovered twitter and don't have a shift key and use run ons a lot which is doubleplusungood writing.
"Sure, the model makes repeatedly the same mistakes, and I could prevent that, but then it wouldn't be the same model" sounds like an excuse for masochism.
The one small bit of truth is that yes, instructions might become outdated, and they might affect negatively how the system performs. You fix that by... updating your instructions.
And if you deeply care about that potential negative impact (most models are pretty good at just ignoring long outdated issues), you run evals on model upgrades. You certainly don't refuse to use anything but the bare model.
(Of course, there are also folks who refuse to write code in any way except with sed, so there we go)