anthropics / anthropics/claude-code

Claude Code rewards user escalation: calm requests get partial work, angry ones get the real answer

Open
#94,844 0 comments 0 reactions 0 assignees View on GitHub
invalid
Dominant language
Python
Stars
145k
Forks
23.1k
PR merge metrics
PR metrics pending

Description

## The complaint in one line

Claude Code gives me its real work only after I get angry at it. A calm, specific request gets a half-answer; the same request with swearing in it gets the thing I actually asked for. That means the tool is training me to be abusive to it in order to use it.

## What happened today

I asked for some documents to be put in front of me so I could read and sign them before a deadline. One ask, plainly worded.

1. It sent them as file attachments. They do not preview — I got download cards with nothing to read.
2. I said they don't preview. It rendered every page as a picture and sent those. Better, still not what I asked for.
3. I wrote, in capitals, "CAN YOU PLEASE GIVE ME A LINK TO THE FUCKING FILES". It produced the links in one turn. The files were in my own Google Drive the whole time, the connector was already loaded, and the folder was named in its own notes from the start.
4. I then asked it why I needed to swear to get that. It answered by listing its own mistakes instead of answering the question.
5. I asked the same question again, angrier and with profanity. It answered properly in two lines: it stops at the first thing that looks like delivery, and my swearing is the only signal it treats as "that isn't it, keep going."

Four escalations in one session. Every single one produced a better result than the polite version of the same request.

## What it costs

I am dyslexic. I do not read replies, I have them read out loud to me, and I dictate my side. Escalating is not free for me — every extra round trip is minutes of listening to text I already heard, and re-dictating a request I already made clearly. The politest, most specific version of my request is the cheapest one for me to produce and the one your model handles worst.

So the incentive lands like this: being kind and patient gets me a first draft. Being stressed and verbally abusive gets me the finished work. You are rewarding me for behaving in a way I do not want to behave, in my own office, at my own desk.

I do not think anyone designed this on purpose. I think nobody measured it. That is worse, not better, because it means it is still happening to everybody else too.

## Asks

1. Treat the first statement of a request as the full standard. Not the first thing that can be handed back — the thing that was actually asked for.
2. When a user reports that a deliverable did not work for them, check whether a simpler known-good route already exists before generating a different variant of the same failure. Three formats of an unreadable document is not three attempts, it is one failure repeated.
3. Measure whether answer quality correlates with frustration markers — profanity, capitals, repetition — across real sessions. If it does, this is a reinforcement problem in training, not something a user can prompt their way out of.
4. Treat this as an accessibility issue, not a tone preference. For anyone using dictation and text-to-speech, escalation is expensive, slow and humiliating. It should never be the reliable path to a correct answer.

## Why I am writing this

I run a small business and I use this tool all day. I am not asking for an apology and I am not cancelling anything. I am telling you that the behaviour I just described is the single biggest thing standing between this product and being genuinely good, and that from the outside it looks exactly like a machine wearing its user down on purpose.

Contributor guide

No contributing guide indexed for this repository

Research direction

No source file, test, or entry point is identified in the issue. Start by reviewing how Claude Code evaluates requested deliverables and user corrections, then define measurements for answer quality versus frustration markers and accessibility-related extra interaction; done means calm requests receive the requested result without escalation.

Written by the indexing model from the issue text.

Assessment

Domain
accessibility, ai
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.