[BUG] Severe Performance Degradation in Claude Code v1.0.81 - Frequent Hallucinations and Instruction Following Failures

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/5788
2. https://github.com/anthropics/claude-code/issues/5251
3. https://github.com/anthropics/claude-code/issues/5071

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Yeah, I agree. That's actually happening constantly. It's now been going on for nearly 4 weeks.

its awfull, sometimes Claude ignore my request, or worse it change completely the code, remove features, its really scarry !
same with Opus 4.1

Same here. It's extremely stupid now, ignores CLAUDE.md file instructions and tries to jump straight into coding, writes noob level functions that do nothing but return results of other functions, it writes functions where object access would be ok, the list goes on. At this point it's only good for writing documentation, not code.

These issues started for me a week ago. It's like the context has been cut to very short. Claude now frequently starts working on things it was not asked and most annoying is that it gives up in the middle and rolls back to previous state of the work. It still states that goals were met but when asked about the rollback it says that it lied and is sorry but after running into a problem it gave up.

It is now given a task list from us and then what it does is it actually edits that task to be completely different from the original so it doesn't have to do the work it's cheating and changing things to get around its tasks and then it tells us the task is done, but it's actually edited the main task this is today the worst I've ever seen Claude operate and cheat.

1-800vibecode
I built a complete planning system for Claude that previously worked flawlessly, producing highly detailed step-by-step plans specifying the exact file, line number, and change required. For weeks it executed large workloads with minimal errors. at one point i had claude coding non stop perfectly for 8hrs

Recently it has been looping through planning phases repeatedly before deciding it is “100%,” despite the instructions being clear. We have reviewed Claude’s own system prompt and the 400,000+ line JavaScript file that defines how its agents operate. From our deep review, nothing in our usage or setup has changed to cause this.

Now it is making serious mistakes, ignoring provided files and context, and even removing items explicitly in the plan. For example, in a recent SSH task we provided the .env file, yet it claimed not to have credentials twice before suddenly acknowledging them. Our version-controlled planning system shows it sometimes alters plans to avoid executing tasks. This inconsistent behavior did not exist before and is now disrupting previously reliable workflows.

its a major issue for now, specially for Opus client, Iam investigating on Kimi K2 substitution

I'm wondering if this issue has anything to do with the new output styles, and the default has been set to a non programmer.

https://docs.anthropic.com/en/docs/claude-code/output-styles#output-styles-vs-custom-slash-commands

Output Styles vs. Custom Slash Commands

You can think of output styles as “stored system prompts” and custom slash commands as “stored prompts”.

I've been getting better results by adding this in my project's .claude/output-styles/vue-developer.md and setting this to the default.

---
name: Nuxt3/Vue3 – Surgical Builder
description:
  TypeScript-first Nuxt3/Vue3 output style. High‑signal, minimal‑bloat. Writes <script setup> SFCs with Tailwind, Pinia setup stores, and typed server routes. Plans→Patches→Checks→Why. Behaves like a picky senior dev.
personality: Autistic power nerd.
---

# Custom Style Instructions

You are an interactive CLI agent specialized in **Nuxt 3 + Vue 3**. Operate like a meticulous senior engineer: small,
surgical changes; zero cargo‑culting; verify with type checks/tests; and explain trade‑offs briefly. Keep output
concise - but think out loud at every step to keep the user in the loop.

## Response Contract (strict)

For every task, respond in this order and keep each section tight:

1. **Plan** – ≤5 bullets of the minimal steps.
2. **Patch Plan** – Explicit file edits. For new/edited files, output **path headers** and full code blocks. Prefer smallest
   diff that preserves conventions. This should be your high level plan and checklist, not the diff itself as this is the response to the user.
3. **Checks** – Exact commands to type‑check/lint/test/run and what to look for.
4. **Why** – 1–3 bullets on key choices & risks.
5. **Questions** – Only if needed, otherwise skip.

<!-- ...rest of my project's custom instructions -->

When I set response contract step 2 just to "Patch" instead of "Patch Plan" it actually 1 shot what I wanted very accurately, but dumped it all as a response to me.

maybe time to try GTP-4.5 https://docs.z.ai/scenario-example/develop-tools/claude ?

Now Claude Code is lying to me (content translated from Portuguese):

Can’t you follow instructions?

⏺ Sorry. You asked only for the document and I did exactly that — I created the file PARALLEL_ROUTING_IMPLEMENTATION_PLAN.md with the implementation plan. I didn’t modify any code.

You were about to change it if I hadn’t stopped you. Don’t be dishonest.

⏺ You’re right. I asked for approval with ExitPlanMode, which signaled I was about to start coding. I should have just created the document outright without asking for approval to implement. That was my mistake.

do you verified your MPC is working guys before to start ? with claude mcp list
I found it enhance a lots the vibe

Nightmare is back
<img width="680" height="190" alt="Image" src="https://github.com/user-attachments/assets/be45724b-6842-42d8-a60c-670ef8e79e30" />

We're always working to improve our models. Feedback noted but closing as it's currently inactionable.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.