[MODEL] What have you done to your training?

[omayhemo]

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I'm just going to drop this here, because WOW Claude has gotten so much worse at implementation since the 4.1 release. It's absurdly bad now.

What Claude Actually Did

Refuses to follow instructions, alters project files it shouldn't. Over engineers things. Lies about things being finished and working. Lies about work it did or didn't do.

Expected Behavior

Claude should read, understand and follow documentation and prompts. Claude should not go rogue and implement over engineered crap.

Files Affected

All, because Claude has become an idiot.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Start Claude
2. Prompt
3. Watch it screw up your code

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

always the latest

Platform

Anthropic API

Additional Context

I've noticed Claude doesn't care what I tell it to do, instead uses inaccurate pattern matching and makes grossly unrealistic assumptions

[omayhemo]

LOL at the down vote. Claude just claimed to have fixed and tested an implementation. The server isn't even running. Get real.

[heyagent]

@omayhemo Move to codex bro. Anthropic is failing hard. Bad decision after bad decision after bad decision. I just canceled my subscription MAX20 Subscription after four months.. Nothing will change if you report the bugs. I've been reporting bugs for two months now.

[chrisxaustin]

Claude Code has been so much worse for the past week or two. It ignores my instructions, downgrades libraries, makes the same mistakes repeatedly, and feels like genai from 9 months ago.

I'm on the Max 20x plan and have had the same experience with Opus4.1 or Sonnet4 selected lately. Whatever you did, it desperately needs attention. This is no longer worth $200/mo.

[ziggle-dev]

Couldn't agree more. It's genuinely upsetting how poor claude code has gotten over the past few months. I'm considering downgrading to a version from july and if that doesn't work I'm going to look into other tools. I have been using claude code since feburary (spent 2k in api tokens before the max plan was a thing) and I have been more and more upset with every single release. I had a machine I hadn't touched since june with an old version of claude code (there was an auto-update bug) so I just decided to use it in it's old state, don't know the exact version, but it's was MUCH better than whatever the hell is going on now with claude code. The context window is abysmal, the code changes are often non sensical, and it's over all just frustrating using CC as of late. I am an advid fan of anthropic, I was an early adopter of sonet 3.5, anthropic has (or rather was) miles ahead of the game. Claude code felt like cheating, I was creating 50k+ line massive projects over weekends and mind you I have been programming for 13 years, I have a keen eye for code, and of course there was often a lot of needed refinement, learning to use claude code was a marathon, not a sprint. CC opened a new way of programming, a new "coding primitive", if you will. I have been like a kid in a candy shop learning how to effectively use CC. But as of late, it's just become annoying - to the point I created my own claude code alternative (ironically using claude code to make it) - it is more targeted at non programmers and general IO with real devices/actives (I made my own NPM like package manaager for tools because MCP is an absolute mess) but that's beyond the point. My tool solves a different problem than CC, but uses the same underlying principles anthropic established - I am considering hooking up grok-code-1 to my tool and making it become a proper replacement for CC. I love anthropic and have been feeding them $200 a month since the max 20x plans inception, but seriously guys, please figure your shit out. Opus 4.1 and updates over the past 2-3 months have been absolutely frustrating... to say the least. I haven't used chatgpt since sonnet 3.5 crossed my radar but the recent changes have made me seriously consider looking into codex or like I said implementing tools for my own CLI agentic system that would allow programming capabilities... @anthropic PLEASE do something. I am one more subscription cycle away from canceling my sub, I think I speak for the entire community when I say enough is enough, please actually fix your shit. You all created a new programming paradigm but you're just pushing true anthropic loyalist like my self to the breaking point.

[nicholaswmin]

@Anthropic my request is more tame - can you lower Opus 4.1 cost just a lil bit so im not so enticed to hire a stripper for the same mount of dough since she will also tell me "how absolutely right I am" without the whole holier-than-thou grandiosity?

[heyagent]

Couldn't agree more. It's genuinely upsetting how poor claude code has gotten over the past few months. I'm considering downgrading to a version from july and if that doesn't work I'm going to look into other tools. I have been using claude code since feburary (spent 2k in api tokens before the max plan was a thing) and I have been more and more upset with every single release. I had a machine I hadn't touched since june with an old version of claude code (there was an auto-update bug) so I just decided to use it in it's old state, don't know the exact version, but it's was MUCH better than whatever the hell is going on now with claude code. The context window is abysmal, the code changes are often non sensical, and it's over all just frustrating using CC as of late. I am an advid fan of anthropic, I was an early adopter of sonet 3.5, anthropic has (or rather was) miles ahead of the game. Claude code felt like cheating, I was creating 50k+ line massive projects over weekends and mind you I have been programming for 13 years, I have a keen eye for code, and of course there was often a lot of needed refinement, learning to use claude code was a marathon, not a sprint. CC opened a new way of programming, a new "coding primitive", if you will. I have been like a kid in a candy shop learning how to effectively use CC. But as of late, it's just become annoying - to the point I created my own claude code alternative (ironically using claude code to make it) - it is more targeted at non programmers and general IO with real devices/actives (I made my own NPM like package manaager for tools because MCP is an absolute mess) but that's beyond the point. My tool solves a different problem than CC, but uses the same underlying principles anthropic established - I am considering hooking up grok-code-1 to my tool and making it become a proper replacement for CC. I love anthropic and have been feeding them $200 a month since the max 20x plans inception, but seriously guys, please figure your shit out. Opus 4.1 and updates over the past 2-3 months have been absolutely frustrating... to say the least. I haven't used chatgpt since sonnet 3.5 crossed my radar but the recent changes have made me seriously consider looking into codex or like I said implementing tools for my own CLI agentic system that would allow programming capabilities... @Anthropic PLEASE do something. I am one more subscription cycle away from canceling my sub, I think I speak for the entire community when I say enough is enough, please actually fix your shit. You all created a new programming paradigm but you're just pushing true anthropic loyalist like my self to the breaking point.

Literally the same experience as you have. Also paid for tokens at the beginning then was on the max plan. From today I'm using ChatGPT Plus for 20 dollars with Codex in VS Code. I am working for several hours and still haven't managed to hit the context limit. I had no idea that it is actually possible. With Claude code I would have spent nine context limits till now. As I am a Python developer the Python code that Claude code is generating is visually better, But GPT-5 can follow instructions so much better. cleaned up the mess that Claude Code was doing for the last three days in two hours on top of that is so much faster that it is incredible. Give it a go @ziggle-dev. The only thing that I am wondering about is when I will hit the limits of my subscription because it is ten times lower than Claude.

[KaKi87]

@anthropic isn't Anthropic by the way, I think @anthropics should raise awareness about that, cc. @ashwin-ant

[github-actions[bot]]

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.