Universal Claude.md – cut Claude output tokens

This doesn't stop it from reasoning before answering. This only affects the user-facing output, not the reasoning tokens. It has already reasoned by the time it shows the answer, and it just shows the answer above any explanation.

Behavior built on top of years and years of experience.

And the problem with AI is that unless you explicitly 'prompt' for certain behavior you're only defining the end result. The inside becomes a black box.

Re- the Unicode chars that are a major PITA when they're used when they shouldn't, there's a problem with Claude Code CLI: there's a mismatch between what the model (say Sonnet) thinks he's outputting (which he's actually is) and what the user sees at the terminal.

I'm pretty sure it's due to the Rube-Goldberg heavy machinery that they decided to use, where they first render the response in a headless browser, then in real-time convert it back to text mode.

I don't know if there's a setting to not have that insane behavior kicking in: it's non-sensical that what the user gets to see is not what the model did output, while at the same time having the model "thinking" the user is getting the proper output.

If you ask to append all it's messages (to the user) to a file, you can see, say, perfectly fine ASCII tables neatly indented in all their ASCII glory and then... Fucked up Unicode monstrosity in the Claude Code CLI terminal. Due to whatever mad conversion that happened automatically: but worse, the model has zero idea these automated conversions are happening.

I don't know if there are options for that but it sure as heck ain't intuitive to find.

And it's really problematic when you need to dig into an issue and actually discuss with "the thing".

Anyway, time for a rant... I'm paying my subscription but overall working with these tools feels like driving at 200 mph on the highway and bumping into the guardrails left and right every second to then, eventually, crash the car into the building where you're supposed to go.

It "works", for some definition of "working".

The number of errors these things confidently make is through the roof. And people believe that having them figure the error themselves for trivial stuff is somehow a sane way to operate.

They're basically saying: "Oh no it's not a problem that it's telling me this error message is because of a dependency mismatch between two libraries while it's actually a logic error, because in the end after x pass where it's going to say it's actually because of that other thing --oh wait no because of that fourth thing-- it'll actually figure out the error and correct it".

"Because it's agentic", so it's oh-so-intelligent.

When it's actually trying the most completely dumbfucktarded things in the most crazy way possible to solve issues.

I won't get started on me pasting a test case showing that the code it wrote is failing for it to answer me: "Oh but that's a behavioral problem, not a logic problem". That thing is distorting words to try to not lose face. It's wild.

I may cancel my subscription and wait two or three more releases for these models and the tooling around them to get better before jumping back in.

Btw if they're so good, why are the tools so sucky: how comes they haven't written yet amazing tooling to deal with all their idiosynchrasies?

We're literally talking about TFA which wrote "Unicode characters that break parsers" (and I've noticed the exact same when trying to debug agentic thinking loops).

That's at the level of mediocrity of output from these tools (or proprietary wrappers around these tools we don't control) that we are atm.

I know, I know: "I'm doing it wrong because I'm not a prompt engineer" and "I'm not agentic enough" and "I don't have enough skills to write skills". But you're only fooling yourself.

The goal here seems to be removing low-value output; e.g., sycophancy, prompt restatement, formatting noise, etc., which is different than suppressing useful reasoning. In that case shorter outputs do not necessarily mean worse answers.

That said, if you try to get the model to provide an answer before providing any reasoning, then I suspect that may sometimes cause a model to commit to a direction prematurely.