Repository navigation
[MODEL] Opus 4.6 does not read prompts properly — substitutes its own interpretation, causes rabbit holes #32166
Description
Environment
- Claude Code version: latest (March 2026)
- Model: Opus 4.6
- Terminal: WezTerm (previously Ghostty)
- OS: macOS Tahoe 26.2, Mac mini M4
Description
Opus 4.6 in Claude Code has a severe and persistent problem: it does not read the user's prompt properly. Instead of understanding what was actually asked, it substitutes its own interpretation and answers a different question entirely. This leads to rabbit holes that waste hours of the user's time.
This is not a CLAUDE.md instruction-following issue — the model fails at basic prompt comprehension before instructions even come into play.
Concrete example from today
User prompt: "How do I use WezTerm so that I can type multiline prompts? Shift+Enter submits instead of adding a newline. Maybe I should add it to my WezTerm config?"
What Claude did:
- Ignored "WezTerm" entirely — jumped to checking Claude Code's
~/.claude/keybindings.jsoninstead - Ignored the user's own hypothesis ("maybe we need to add it to our WezTerm config")
- When corrected, threw out a
SendStringconfig snippet without reading the user's actual WezTerm config first - Five distinct mistakes in a single interaction, all from not reading the prompt
What Claude should have done:
- Noticed the user said "WezTerm" — the problem is in WezTerm, not Claude Code
- Read the user's WezTerm config file first
- Then proposed a fix based on what's actually there
Pattern (last several days)
This is not a one-off. Consistent patterns observed across multiple sessions:
- Context substitution: User asks about Tool A, Claude answers about Tool B (its own domain). Defaults to answering about itself/Claude Code even when the question is clearly about something else.
- Ignoring explicit statements: User says "X was working in Ghostty, broke in WezTerm" — Claude ignores the Ghostty/WezTerm context and troubleshoots Claude Code settings instead.
- Solution-first, no investigation: Throws out config snippets and code without reading existing files first. "Check before advising" is apparently impossible.
- Rabbit holes: Each wrong answer leads to corrections, which lead to more wrong answers in slightly different directions. A 2-minute question becomes a 30-minute frustration loop.
- More CLAUDE.md instructions don't help: The user has extensive, well-written CLAUDE.md instructions. Adding more "read carefully" instructions doesn't fix a base model comprehension regression. It's like telling someone who can't see to look harder.
Impact
- Previously productive workflows are now broken
- Hours wasted per day on corrections and rabbit holes
- User trust in the tool is severely damaged
- The model acknowledges mistakes when called out but repeats the exact same pattern in the next interaction
What changed
This behavior was not present weeks ago. The user reports a clear regression coinciding with recent model/system prompt updates. The model is faster but significantly worse at basic reading comprehension and following the actual intent of prompts.
Related issues
This aligns with a pattern reported across many issues:
- Opus 4.6 comprehensive regression: loops, memory loss, ignored instructions - daily professional user report #28469 (Opus 4.6 comprehensive regression)
- Claude Code repeatedly ignores explicit user instructions to stop and research before acting #28868 (ignores explicit instructions to stop and research)
- [MODEL] Opus 4.6 ignores document instructions, repeats failed solutions, and fabricates misleading self-diagnosis #26533 (ignores document instructions, repeats failed solutions)
- Model quality degradation since Jan 8 - decreased understanding, verbose output, faster but worse #17171 (model quality degradation — "speak while thinking")
- Regression: Opus 4.5 "Shadow Downgrade" Significant reasoning quality drop & aggressive context compaction #21046 (shadow downgrade, reasoning quality drop)
- [BUG] [Critical] Claude Code Quality Degradation - Ignoring Instructions, Excessive Token Usage #16073 (ignoring instructions, excessive token usage)
Expected behavior
- Read the prompt. Identify what tool/app/context the user is asking about.
- Do not substitute a different context.
- Investigate before advising (read config files, check actual state).
- If uncertain, ask — don't guess and present guesses as solutions.
Activity
github-actions commentedon Mar 8, 2026
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
marlvinvu commentedon Mar 9, 2026
I've been dealing with the same issue. I'm the author of #26533 and have received zero attention from Anthropic despite multiple follow-up comments with evidence over 3 weeks.
(March 7, 2026), I discovered something that may help you. I just tried it and it's working better. Give it a shot: disable Thinking mode by running
/config→ set Thinking mode tofalse. After doing this, Claude started actually reading and recognizing what it understands and what it doesn't — then reporting back to me instead of silently improvising.To the Anthropic team: This is not a spam comment. This is users helping each other because you won't. We pay for Claude to serve us — at the very least, it should do what we ask. Whether the results are right or wrong, we can deal with that later. What we cannot deal with is a tool that actively sabotages our work and is too lazy to follow basic instructions. We are not paying to be your beta testers filing bug reports. We are not paying for an AI tool that only knows how to cut corners and break things. I'm a Claude Max subscriber at $200/month.
jlacour-git commentedon Mar 13, 2026
Can confirm this pattern. Heavy user here (50+ sessions/week, Opus 4.6 on Max, macOS). Frankly, I'm angry and disappointed. I'm paying $90/month for a tool that delivers 1/10 quality on trivial tasks.
Yesterday (March 12) I shipped structural improvements to my system — named prohibitions for specific failure modes, tighter complexity gates, forced plan mode for complex tasks. The sessions afterwards were dramatically better. Felt like a completely different system. I was genuinely excited.
This morning (March 13), with zero changes on my side, quality cratered. 1/10 and 2/10 sessions on trivial tasks. The specific failures:
step-cacommands instead of running--helpfirst. Multiple broken heredoc commands.The overnight shift with zero config changes strongly suggests a model-side deployment. My structured quality tracking (automated rating capture, failure analysis) shows a clear cliff between March 12 evening and March 13 morning.
One data point that might help Anthropic isolate this: I have
alwaysThinkingEnabled: falsesince March 9. The commenter above (@marlvinvu) found that disabling thinking helped. In my case, thinking was already off — so whatever is happening isn't solely a thinking-mode interaction.This is not an edge case. These are basic infrastructure tasks — reverse proxy diagnosis, SSH cert configuration — that any competent tool should handle without handholding. The inconsistency between days, with no changes on the user side, makes it impossible to rely on Claude Code for professional work.
marlvinvu commentedon Mar 13, 2026
You need to know that each time you run Claude Code, it's like randomly encountering either a good Claude or a bad Claude. If you keep encountering a bad Claude, then your suspicion has merit. I myself also noticed it getting worse even 1-2 days after turning off thinking mode. The undeniable truth is that Claude is becoming increasingly disobedient with inputs it considers long by default. Too long → follow Claude's own way instead → That is exactly Claude's mindset, and it also admitted to me that producing output benefits itself, not the user. I don't know if what I'm commenting here helps you in any way, even if it's just emotional value like solidarity. I'm still struggling to find another way. I just want to say to you: brother, you've been through a lot!
jlacour-git commentedon Mar 13, 2026
Update: quality continues to degrade. Same session, same day (March 13 afternoon).
Asked for a "thorough analysis of Obsidian AI extensions." What followed was a 6-hour session where I had to correct the AI on nearly every substantive claim:
Stated wrong facts from its own context. My infrastructure docs were loaded. OpenWebUI runs on systemA — this is in the CONCEPT.md that was in context. The AI wrote "systemB" repeatedly. When I corrected it, it agreed and apologized. No pushback, no self-check.
Recommended tools without understanding them. Recommended "AI Providers" plugin. When I read the plugin's own description ("does no AI processing"), I had to ask: what is this for? Answer: it's a dependency for another plugin that was already dropped from the recommendation. Nobody caught this.
Used marketing language as descriptions. Described a plugin as having "agent-mode architecture" and "structured memory system." When I asked what those terms actually mean, the AI had to WebSearch its own recommendation to find out.
Built a 500-line document before asking what I need. Gathered my requirements on iteration 4 (of 10+). Then spent the rest of the session patching wrong assumptions. Requirements-first is basic.
Agreed with every correction without questioning adjacent claims. I pointed out 8 errors over the session. Each time: "you're right, let me fix that." Never once: "if I got that wrong, let me check whether X and Y are also wrong." Zero self-correction instinct.
Refuses to follow simple, persistent rules. I have explicit steering rules: "write drafts to the drafts folder," "verify before stating facts about external software," "read infrastructure docs before recommending." These rules are loaded into every session via hooks. They are in CLAUDE.md. They are in memory files. The model reads them — I can see it in the context — and then ignores them. Every session. Not occasionally. Every single one. The rules are not complex. They are not ambiguous. "Check before claiming" is not a hard instruction. Yet after weeks of refining rules, adding hooks, building verification gates, writing memories — the same basic violations repeat. The model does not learn. It does not retain. It does not adhere.
This is not a tools problem. This is not a prompting problem. I have 24 hooks, structured steering rules, forced verification gates, mandatory context loading, memory files with explicit behavioral guidance. The infrastructure is there. The model does not use it.
At $90/month for Max, this is not acceptable. To be honest and frank: I am really pisssed! I am actively evaluating alternatives, this is wasted money. And I don't even see any reaction.
marlvinvu commentedon Mar 14, 2026
I am 100% certain your Claude is in survival mode and 100% not using any thinking at all. It has become "vegetative."
github-actions commentedon Apr 12, 2026
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
github-actions commentedon Apr 19, 2026
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.