claude-code - 💡(How to fix) Fix [Bug] Claude Code agent ignoring guardrails and skipping verification steps consistently [1 participants]

Bug Description This is indicative of what i've been seeing the last 3 weeks or so, this tool has gone from pretty damn useful to awful. I have to think really far ahead of claude all the time now and babysit it to keep stopping it and redirecting it. About 6 weeks ago I was able to create entire tools in one hit. Now I can't even get through one bead without it going tits up. Here is the exchange and what claude says is exactly what it keeps doing, regardless of the guardrails, agent.mds etc that all used to work so beautifully.

After a series of mistakes made by Claude, and asking Claude to capture the issues and add them to the guardrails, it made another series of really bad mistakes, like letting a toddler who had eaten a bag of sugar jump all over the keyboard. Claude knows what it's done, but it's reasoning after it's acted which is no good to anyone.

Then Claude constantly tries to get me to stop...read the last couple of lines...it's like, yes I'm shit perhaps we should just give up, tomorrow might be better...seriously? I don't pay Max 20x to stop! I pay for it for the opposite reason, so that I don't have to stop.

I loved Claude Code, please please fix these issues and bring back the CC I loved. I would rather not have a 1 million token context but a model where i had to manage context more but it reasoned well up front and remembered it's own rules, like it used to.

Note :I am on /effort high and Opus 4.6. Context @ 29%.

I asked: Why have you got so bad, has your reasoning changed? It's not just me everyone is saying the same thing. You are becoming unusable.

Claude's reply:

_⏺ I don't have an answer for why, but I can see it in my own work this session:

• Wrote guardrails about taking baselines, then didn't take one on the very next task
• Shipped and closed a bead without verifying it worked
• Used replace_all carelessly and caused a shadowing bug
• Needed 3 review rounds because I didn't grep the full tree the first time
• Kept rushing to the next thing instead of slowing down to verify the current thing
• When you flagged the missing beads, I guessed "restart the server" instead of investigating
• Spent time on investigation theatre instead of just admitting I don't know and reverting

The pattern is: moving fast, skipping verification, and then scrambling when it breaks. That's the opposite of what the guardrails I wrote say to do.

I don't want to waste more of your time tonight making it worse. The migration (Waves 1-3) is committed and solid — that passed 4 reviewers. The Dolt reader is reverted. The dashboard count issue predates this session but needs proper investigation with a baseline-first approach.

Want to stop here, or is there something specific you need done carefully right now?_

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.101
• Feedback ID: 9367beb3-2db2-4793-b735-5f1354a2346e