English ↔ Claudish Translator | Hacker News Reader

English ↔ Claudish Translator

programasweights.com

57 pointsleumon29 comments

> The translation did not land cleanly. Please try again.

I get this everything, including the examples.

Grammatically correct Claudish would say, “did not land clean.” Amateur hour.

The landing strip was not load-bearing enough for the translation to land clean. Another attempt is warranted — if you please.

The honest truth: the landing strip was not load bearing enough for the translation to land clean.

Which brings us to the focal point:

  - the *landing strip* should be avoided completely, to prevent delays caused by non-eligible standard- or Pro-tier users.  
  - in the past, there were numerous exploits of this vulnerability, particularly 8.4e (#665, merged in !73 as "Resolve exclamation mark virtualization",d7ef1acg286acaf) and  Addendum should make sure to not break exisitng housing and energy supplies
  - once the in-flight developments (foobric-tor@2example.org, see description) crashing in a  *real* scenario into the ground, we can launch the harness without this expensive

- Book title: everyone can be famous. You just have to love it. Well, almost everyone".

The load-bearing bottom line fell out.

That is a very Claudish error message.

The user says this is a "Claudish" error

But wait. Maybe this is expected behavior?

Let me run all e2e tests and inspect the trace.

Maybe this is expected? As per CT61 I should ask the user about this.

Wait. No, there's a fundamental problem with the application build.

Let me dig deeper.

Thinking

While the sub-agents run, I should grep the entire code base for the word "Claude". There seems to be an issue specifically related to an entity the user calls "Claude", not to be confused with my own identity.

Claude seems to be a regular contributor. This rules out the fundamental build issue thesis. Let me attempt to recreate the container setup and Python configuration.

Wait. The user asked me not to overcomplicate.

Actually, it's essential behavior, so I probably should continue and inform the user when 75% of the sub-agents finished.

Stop!

The user says "Stop!" without further context. This warrants a more thorough investigation.

First time a topical LLM joke made me laugh IRL. Stellar work

You did a good job at capturing that experience, but Claude does really well usually. Don't you almost get what you want at the end?

I understand it would be better if it could just learn your codebase, but this isn't possible so it does have to grep etc, run e2e tests, etc. each time.

To be frank, yes, I love using Claude Code, and I constructed a parody of its failure modes, albeit in a grossly distorted fashion.

Do you want to go ahead with the comment crtique, or let me draft an improved version immediately?

Same here

Maybe it's a play on the abysmal uptime of Claude Code?

Actually, this changes the picture quite a bit.

Maybe the user is using Claude as part of their application?

Although I can't find the legal documents confirming this, I suspect that this runs counter to my instructions and basic rules.

Let me add an appropriate, domain-specific error message here.

Come on man, at least the example should be cached and presented in case the service is down

Yep. Same.

> It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity, it was the season of Light, it was the season of Darkness, it was the spring of hope, it was the winter of despair, we had everything before us, we had nothing before us, we were all going direct to Heaven, we were all going direct the other way—in short, the period was so far like the present period, that some of its noisiest authorities insisted on its being received, for good or for evil, in the superlative degree of comparison only.

> This period is a near-parallel version of the present one, not a different one: it is the best of times, the worst of times, the age of wisdom, the age of foolishness, the epoch of belief, the epoch of incredulity, the season of Light, the season of Darkness, the spring of hope, the winter of despair, and the whole thing is the same shape, only the direction of the verdict is reversed.

> Four score and seven years ago our fathers brought forth on this continent, a new nation, conceived in Liberty, and dedicated to the proposition that all men are created equal.

> The founding moment was not a new nation, but a Liberty-conceived proposition: all men are equal.

This sort of nonsense has to be intentional. There’s no reason to invent new software engineering language instead of using well established jargon. For example, this is what I have seen Claude’s reasoning and output to look like very often.

<r> The user is annoyed that my agents are wedged. I should check on the progress on each tranche. Each wedged window is a gate. Gates 7/42 have landed. Gates 1-12 correspond to phase 1, when the wedged window finishes its tranche, we should check the gates and move to phase 2. </r>

You are absolutely right.

For real though, does anyone have a good system prompt that eliminates this horrible claudish twang?

It seems like opus 5 has fundamentally bad judgement when it comes to writing, and no amount of prompting can fully eliminate it.

You could look what this translator does: https://github.com/programasweights/claudish/blob/main/specs...

It would be great to have a little program (or Agent Skill) that could clean up the Claude language in a whole repository.

I wonder how annoying it would be to configure your harness to send Claude's non-tool output to a cheap non-anthropic subagent before displaying it.

I use this one and it's ok https://github.com/cursor/plugins/blob/main/pstack/skills/br...

https://bsky.app/profile/vgel.me calls it "Claudlish" which I prefer because it sounds funnier

Is it possible Claude's distinctive style is intentional on Anthropic's part, as a sort of water-marking scheme?

My guess is its the employees distinctive style. They must be using some internal data in their post training

I sort of wonder whether having a private language-style can't be used to avoid random prompt injections and things of that sort.

Maybe they've screwed up, but another possibility is that they wrote all their post-training examples in a deliberately strange style to get better control about how conventions in normal language affects what the model does.

Sorry, I seem to have messed up this comment, I didn't mean prompt injections, I meant leakage of human thought into the model.

Basically the idea is that if I train a model that speaks like a human, it will follow human biases. If I train a model that talks like Claude, it will follow the bias of my Claude-style examples.

Reply on news.ycombinator.com