claude-code - 💡(How to fix) Fix Recurring output degradation across Sonnet 4.6, Opus 4.6, Opus 4.7, and Haiku 4.5 — April fix did not hold

1.	Began the front-end design task but ignored its own checklist entirely
2.	Made unauthorized changes to back-end files that were never part of the request
3.	In doing so, broke existing back-end functionality that was previously working
4.	When confronted with the damage it caused, attempted to fix it multiple times but failed on every attempt
5.	After exhausting all attempts, the task was handed off to Codex, which resolved the back-end issue in a single prompt

1.	Received a clear, specific request to find GitHub repositories covering certain topics
2.	Returned results from completely unrelated repositories that did not match the request
3.	When corrected and given additional clarification, repeated the same mistake across multiple follow-up attempts
4.	Never once returned results relevant to the original request despite repeated clear instructions

Back-end files were modified without authorization during a front-end design task. Exact file names are not recalled, but the changes broke existing back-end functionality that was previously working. No back-end files should have been accessed or modified at any point during this task.

1.	Give Claude Code a front-end design task with a defined checklist
2.	Observe that the checklist is not followed
3.	Check back-end files after the task — unauthorized modifications will be present
4.	Ask Claude to fix the back-end damage it caused
5.	Observe that it fails to resolve the issue after multiple attempts

1.	Ask Claude to find specific GitHub repositories covering defined topics
2.	Observe that returned results are unrelated to the request
3.	Correct Claude and repeat the request with clear explanations
4.	Observe that irrelevant results continue to be returned across multiple attempts

1.	Open any Claude model (Sonnet 4.6, Opus 4.6, Opus 4.7, Haiku 4.5)
2.	Submit a simple, straightforward task
3.	Observe circular responses, failure to follow instructions, and overall degraded output quality
4.	Compare output quality to pre-March 2026 performance — the regression is consistent and reproducible across models and sessions

Exact conversation logs were not saved, however the following patterns were observed consistently:

Claude Code:

	•	Claude narrated its steps while working on the front-end task but made no mention of why it was modifying back-end files
	•	When asked to fix the back-end damage, Claude acknowledged the issue each time but failed to resolve it, often repeating the same failed approaches across multiple attempts

Claude Online (claude.ai):

	•	When returning wrong GitHub repositories, Claude presented the results confidently with no indication that they were unrelated to the requested topics
	•	When corrected, Claude acknowledged the mistake but continued returning irrelevant results in subsequent attempts

•	Degradation is consistent across multiple sessions, not isolated incidents
•	Quality issues affect all tested models: Sonnet 4.6, Opus 4.6, Opus 4.7, and Haiku 4.5
•	Claude Code consistently oversteps its scope, modifying files outside the requested task
•	Claude online consistently returns irrelevant results even after repeated corrections
•	Behavior mirrors the degradation patterns reported and acknowledged by Anthropic in the August–September 2025 and April 23, 2026 postmortems, suggesting the fixes did not fully hold

•	Front-end tasks repeatedly result in unauthorized back-end modifications
•	GitHub repository searches repeatedly return unrelated results regardless of how clearly the request is framed
•	When Claude breaks something, it is unable to self-correct after multiple attempts

•	A competing tool (Codex) resolved the back-end issue Claude caused in a single prompt, highlighting the severity of the regression
•	This report is being filed as part of a broader pattern of quality degradation noticed by the developer community, not as an isolated bug