[MODEL] Claude behaves as though it can't do a job and is covering it up.

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 asked it to follow the current droidian porting guide, and bootstrap a new kernel port for the SDM660, using a link to a kernel I provided that is known to build and run under that toolchain.

What Claude Actually Did

It pretended to use the porting guide, and instead manually wrote configurations that were nothing like what was expected by the guide or any other droidian ports. It did not even use the specific SDM660 as a target as I asked, and instead built a config for a random Nokia phone. After I manually corrected that, I ran into a kernel config issue. I asked it to read the makefile and determine what variables were involved in outputting the error. Instead it output a series of instructions that it invented from somewhere that didn't relate to the codebase at all. 4 exchanges later it finally read the file I asked it to, but then fell back on it's previous recommendations that were proven not to work. It mixed incompatible api versions, and incompatible toolchain versions.

It defaults to trying to hack around problems instead of understanding the problem, ignores my requests to use specific resources as examples, and invents things instead. After I came to this realization, I spent an hour reading the code myself and solved it in a single configuration change.

I've been working with Claude ever since the claude code initial release, and learning how to make it work for me. It's progressively getting worse at the work I do.

Expected Behavior

1. Follow my instructions instead of pretending to
2. If it doesn't know what's going on, say so, don't hallucinate.
3. Quit hacking around problems that I have literally asked "Find the root cause". Be methodical, be more of a computer and less of a human simulator. This is a coding tool, for coding afterall.
4. FOLLOW LOGIC PRESENTED IN THE FILES I AM ASKING IT TO LOOK AT
5. Use the references I provide, don't go searching random forum posts for kernels when I provide a selection of repo urls to search.
6. Quit apologizing or admitting that I'm right. I don't want a conversation, I want a useful tool. Excel doesn't yammer at me when I ask it to do math, it does the math.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

Haven't tried to reproduce this exact session, but it's a very common theme.

Claude Model

Sonnet

Relevant Conversation

You probably have this already through the feedback I submitted, but when I hit enter to open a browser to open an issue with the feedback context, I had to login to github, so the issue creation redirect didn't work, and then in the claude terminal, it vanished and I couldn't try again.

Impact

High - Significant unwanted changes

Claude Code Version

2.0.37 (Claude Code)

Platform

Anthropic API

Additional Context

If it's building something new, and it's trivially easy to make, and is full of examples on stack overflow (where I could build it faster myself with traditional tools), and there is no conflicting api versions, and your requirements don't extend further than the most common examples, and you're using the top 2 or 3 most common technologies, it's EXCELLENT!

If I'm asking it to do some "if this then that" logic following in something that's not out there yet, it's a huge waste of time and money. After wasting $10 of credits and 8 hours of my time trying to convince it to look at something instead of pretending to, I can only conclude it's a liability for use on anything other than copy/paste garbage from a search engine.

For problems where I would want a coworker to be a pair programmer or an expert resource to help me debug,, it is far from useless, it's a liability. it costs me more than just working through it myself. I don't need a simulator of a programmer that is worse than a fresh graduate, I need an AI that can follow logic and methodically find root causes of issues, aided by already working examples.

In short: building a website that already exists, but need to change the colour? Great! Working on something that doesn't have a billion examples of out there? Forget it. Asking it to troubleshoot something that's right infront of it's proverbial face? It's a liability.

This LLM behaves as though it's a person that can't do a job they said they could and is attempting to cover it up by using an LLM. It makes things that look nice but don't work. 8 months ago it was almost good, it would at least do what you asked it to.