[Bug] Claude Code: Persistent Instruction Disregard and Output Quality Degradation

this is tough to deal with and I am considering moving away from Claude Code until this has been resolved and this is really sad I will have to go to the Cursor Ultra package which gives me much less usage. Btw this is not happening with Claude Sonnet through Cursor which I will have to now use. I made a Bug the other day about quality of outputs also and have not gotten any replies. please help me here I am willing to collab or show you but this is bad enough that is it ongoing 3 days at least of low quality outputs due to lack of following direction without threats

Same experience. Its been degrading a lot. To the point where I keep checking which model I am using. It doesn't use its agents, it doesn't read the instructions and it dreams up modules and functions that do not exist. It is difficult to not assume that intelligence is being throttled to be used for other purposes. For the first time in a very long time I am convinced I would have achieved more yesterday if I did not use Claude Code at all. Getting to my desk in the morning knowing that I need to wrestle a todler to get anything done is not a pleasure.

I would love to continue using Claude Code because its the OG and until recently it "just worked".
I am currently considering codex-cli, cursor-cli, augment-code.
All tools I would not consider if Claude Code worked like it did a while back.

Latest update I am working on getting a support rep to help with this

Comprehensive Bug Report Summary for Claude Code Team

GitHub Issue: #8043 - https://github.com/anthropics/claude-code/issues/8043

Issue Title: Critical regression in Claude Code instruction-following and edit functionality

Timeline:

• Long-time Claude Code user (max account subscriber)
• Issue began ~5 days ago
• Started AFTER the announced quality fix deployment - https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues
• Currently running one version back from latest release
• Was working fine just a week ago with only occasional edit failures

Primary Issues:

1. Instruction Non-Compliance:

• Ignores explicit file existence confirmations
• No longer checks if files exist before creating them (even when instructed)
• Disobeys direct commands
• Requires extremely aggressive prompting to get basic compliance
• Quality of responses degrades rapidly over conversation length

1. Edit Functionality Breakdown:

• Edit failures occurring constantly ("left and right")
• "String not found in file" errors persistent
• Having to resort to sed and echo commands as workarounds
• Creating errors all over files when forced to use alternative methods

1. Inconsistent Behavior Across Sessions:

• Running 10 agents simultaneously: ~6 fail on edits, 4 work fine
• One terminal session fails repeatedly, another works without issue
• Not related to simultaneous file editing (user aware of that limitation)

Technical Details:

• Platform: Windows (win32)
• Terminal: windows-terminal
• Version: 1.0.120 (one version back from latest)
• Feedback ID: fa462473-a581-4bb5-8cf1-fd8697143392
• Context: Staying within 200k window, not hitting technical limitations

Specific Bugs:

• /compact fails when run manually but works on auto-compact
• Low quality outputs even with examples provided (which user "almost always" provides)
• Multiple "Request was aborted" errors

Comparison Testing:

• Claude Code: Severe degradation, edit failures, ignores instructions
• Cursor with Claude Sonnet: Works perfectly with same prompts/projects
• Confirms issue is Claude Code-specific, not model-related

Impact:

• Forced to purchase Cursor Ultra subscription as workaround
• Workflow completely disrupted despite paying for max Claude Code account
• GitHub reporting not working effectively for user
• Considering abandoning Claude Code entirely

Latest repeat errors

⎿  API Error: 500 {"type":"error","error":{"type":"api_error","message":"Overloaded"},"request_id":null}

When attempting to use Claude Code again tonight Claude through Cursor still working fine. Worked well for 30 min

I'm coming from VS Code's version of Copilot, but I can also report having new issues with Claude Sonnet 4 failing to listen to instructions in .github/copilot-instructions.md.

Frequently, despite my instructions file being a reasonable ~200 lines, Claude will disregard basic steps it should take. For example, my instructions file has the following steps in it:

## Steps (always)

1. Open or create **PLAN.md** with a comprehensive checklist (all steps/substeps to complete the project).
2. Pick the next unchecked item → **write tests first** (≥3, incl. edge cases).
3. Run `cargo make lint` and fix all issues; run tests; ensure coverage **≥ 80%** via `./scripts/check-coverage.sh`.
4. Implement; re-lint; re-test; update PLAN.md and check off completed items.
5. Run a dedicated build step to ensure the app compiles cleanly.
6. Final code review (security, performance, maintainability, plan adherence); you may use the **changes** tool for this pass.
7. Stage and commit; commit hooks must pass tests & lints. **Never** use `--no-verify`.

Despite these very clear instructions, Claude consistently skips on steps 4 and 6 and sometimes fails to do TDD altogether (ignoring step 2).

I would love to see it work like it did 2 months ago, as this is a bit frustrating. I'm almost getting similar results with Gemini, which I never thought I would say.

As for Claude Code, I have not tried it in the last week or so, so I cannot speak on it in particular. Just wanted to report degraded output from Claude Sonnet 4 in GitHub Copilot.

Another update

Getting alot of - ⎿  Error: File has been unexpectedly modified. Read it again before attempting to write it.

Please see -

● Update(ThothOS\src\app\api\auth\mfa\generate\route.ts)
⎿  Error: File has been unexpectedly modified. Read it again before attempting to write it.

● Read(ThothOS\src\app\api\auth\mfa\generate\route.ts)
⎿  Read 71 lines (ctrl+o to expand)

● Update(ThothOS\src\app\api\auth\mfa\generate\route.ts)
⎿  Error: File has been unexpectedly modified. Read it again before attempting to write it.

● Let me check the next MFA route file.

● Read(ThothOS\src\app\api\auth\mfa\get-secret\route.ts)
⎿  Read 129 lines (ctrl+o to expand)

● Update(ThothOS\src\app\api\auth\mfa\get-secret\route.ts)
⎿  Error: File has been unexpectedly modified. Read it again before attempting to write it.

● Let me check a GraphQL route file instead.

● Read(ThothOS\src\app\api\graphql\accounting\billing\creditcard\customer\route.ts)
⎿  Read 231 lines (ctrl+o to expand)

● I need to add the withTracking import and wrap the POST function.

● Update(ThothOS\src\app\api\graphql\accounting\billing\creditcard\customer\route.ts)
⎿  Error: File has been unexpectedly modified. Read it again before attempting to write it.

● Let me try a different approach. Let me check if the tracking wrapper already exists and see what it looks like.

● Read(ThothOS\src\lib\tracking-wrappers.ts)
⎿  Read 135 lines (ctrl+o to expand)

● Let me check a categories route file.

● Read(ThothOS\src\app\api\graphql\categories\article\administration\route.ts)
⎿  Read 236 lines (ctrl+o to expand)
⎿  Interrupted · What should Claude do instead?

use the orchestrator dude

● I'll launch the api-tracking orchestrator to continue updating the API endpoints.

● api-tracking(Update API endpoints)
⎿  Read(ThothOS\src\app\api\graphql\categories\article\administration\route.ts)
Read 236 lines (ctrl+o to expand)
+4 more tool uses

This is an example of the degradation in performance. Nothing is editing those other files that is actually incorrect it isn't being modified.

It works great sometimes other times its this. Still no response from the Anthropic team on this. I get I am reporting multi issues here but its a representation of the reduced quality. This is causing the model to literally skip work because it cannot edit and assumes it was done in another agent even though it clearly reads the file and it hasn't been editted!

Edit: Email from the Anthropic team not addressing this but instead sending me generic docs. Will update when I have one

Sonnet 4.5 works alright so far. So this is a Sonnet 4.1 and Opus 4.0 issue. I am fine with going to 4.5 if thats the fix so be it. leaving open for a few days to make sure

It would seem to me that you need a better way to monitor your issues in GitHub. Also a single automated email is kind of unacceptable from Bug reports. Not sure what your long term vision is but with the best model in the world for coding I would think you would want to provide better customer service. Maybe you need help I don't want to assume incompetence or negligence - email me if you want to do something 10+ years in network and software engineering - mkgoluba@technologiesunlimited.net

there are some standards you could create with github actions that would help to put priority issues on your radar. matter of fact i know i am not the only one with response concerns the issue is obvious that your github issues are not being touched for weeks at a time. I can help if you want to fix this reach out please.

Leaving this open to prove the point. I expect it will be weeks before I get a response if i get any

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.

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.