Significant quality regression — unauthorized actions and unreliable instruction following

Summary

Claude Code has become significantly less reliable over the past month. It frequently ignores its permission model, regularly disregards configured guidance (CLAUDE.md conventions, custom skills, learnings files, session corrections), and often invents its own approaches without asking. The regression is severe enough that I'm considering not renewing my subscription.

Specific Problems

Unauthorized / Unrequested Actions

• Asked to read a file — Claude edited it instead
• Asked to investigate a bug — jumped straight to "fixing" code before I approved a diagnosis
• Told to stop and wait — continued making changes
• Creates commits without being asked
• Adds unsolicited refactoring, docstrings, and "improvements" to code I didn't ask it to touch
• Invents new implementation patterns and rewrites code without discussing the approach first

These don't happen every time, but they happen often enough that I can't trust the tool to stay within the bounds of what I asked.

Unreliable Instruction Following

I've invested significant time configuring:

• CLAUDE.md conventions — Coding standards, banned patterns, file organization. Claude frequently violates them, sometimes even ones marked "STRICTLY PROHIBITED"
• Custom skills — Domain-specific guidance that gets overlooked
• Learnings files — Documented workarounds from past debugging sessions. Claude often repeats the same mistakes these were written to prevent
• In-session corrections — I correct Claude, it acknowledges, then makes the same mistake again later in the session. Not every time, but far too often

The guidance works sometimes. But the inconsistency is the problem — I never know whether Claude will follow the conventions or go its own way, so I have to verify everything regardless.

Performance / Reasoning Degradation

• Frequently loses track of the task mid-conversation and goes on tangents
• Generates incorrect code confidently, requiring multiple correction rounds
• Gives generic answers instead of reading the actual code
• Makes up function signatures and patterns that don't exist in the codebase
• Fails at tasks it handled correctly a month ago with the same configuration

Workflow Impact

• Tasks that used to take 5 minutes now take 20+ minutes of corrections
• I often spend more time reviewing and reverting than I would writing the code myself

Impact

I'm a paying Max plan subscriber who uses this tool daily for professional work. The current state is that Claude Code is unreliable enough that it's become a net negative on many tasks. When I can't predict whether the tool will follow instructions or go rogue, the oversight cost eats up the productivity gains.

I've put real effort into the configuration system — CLAUDE.md, skills, learnings, memory. When those work, the tool is great. But the hit rate has dropped significantly over the past month, and the unauthorized actions have gotten worse.

I'm seriously considering going back to my pre-AI workflow when my billing cycle ends.

Environment

• Platform: Linux (Debian)
• Claude Code CLI
• Model: Opus
• Subscription: Max plan