Sustained quality regression, 7 Sep – 9 Oct 2026: agent abandons stated task, Romanian dictation leaks stale text, new-chat routing defect, model tier not honored

Sustained quality regression in Claude Code, 7 Sep – 9 Oct 2026 (single-user report)

Reporter: solo user, non-developer by trade (agronomist). Paid subscriber. Uses Claude Code
daily as the primary tool for a hardware project (CAD, firmware, BOM) plus personal
knowledge/logging infrastructure.

Summary: since roughly 7 September 2026 this tool has been unusable for sustained work.
Not "sometimes wrong" — the failures are frequent enough that entire 5-hour session windows get
consumed by the agent's own mistakes rather than by the task. I moved to a competitor for
17–28 September because I could not work here at all. The gap in my logs is not a quiet period;
it is the period I was gone.

What breaks, with dates from my own session transcripts

1. The agent leaves the stated task.
Asked for robot/CAD work; the session goes into the agent's own infrastructure concerns instead
and the window is spent. 12, 13, 14 September, explicitly flagged by me each time; again 9 October.
On 12 Sep: "I wanted to work with you in Blender on our robot, and all we did was you fucking up
my night." On 14 Sep: "it's about the brain and i want the robot — we sorted this yesterday."
This is the single most frequent failure mode, more frequent than any tool or infra bug.

2. Confident contradiction of things I know to be true.
9 and 14 September. The agent asserted, with confidence, that facts I had personally logged were
not logged — on the basis of one weak lookup. "Why do you keep contradicting me on stuff I'm sure
about? What is that confidence you attacked me with if no one told you?" Rule-following
degrades across a session: constraints I stated minutes earlier get dropped.

3. Minor, mechanical tasks fail.
Not just hard problems. Color-matching three parts against reference photos I uploaded had to be
corrected repeatedly (13 Sep). A chassis-geometry regression reappeared after being resolved
(15 Sep). These are the cheapest possible tasks.

4. Romanian dictation is unusable and has been for months.
I dictate because I have dental problems and typing long prompts is slow. Anthropic's dictation
mangles Romanian badly enough that the agent then reasons from garbled input and produces
confabulated output — on 9 September it invented a mechanical part because "șurub și piuliță"
(screw and nut) came through wrong. Siri's keyboard dictation on the same machine handles Romanian
better. 9 Sep, 6 Oct, 9 Oct.

5. Interface defects nobody appears to own.

• Clicking "new chat" in the desktop app routed my prompt into the previously running session

instead (12 Sep). "I nearly broke both my keyboard and my desk chair of how annoying this is."

• Starting a session from the cloud/GitHub path silently created the wrong kind of session and

destroyed a working night (29 Sep, again early Oct).

• These are not model problems. They are the surface the user actually touches, and they are

the ones that cost whole days.

6. Model routing.
Repeatedly found myself on Sonnet when my configured default is Opus (29 Sep). My
~/.claude/settings.json has read "model": "opus[1m]" since then, and it still happens. If this
is intended behaviour under load, say so in the UI; silently serving a different tier than the one
configured, to a paying user, is not acceptable.

Why this report exists

I am one person, not a team with a QA budget. A keyword sweep of my own local session transcripts
between 5 August and 9 October 2026 finds 17 separate days and 79 separate messages in which I
told the agent, in the moment, that it had derailed, contradicted me, ignored a stated rule, or
burned my session. The worst single days: 13 September (5 sessions, 17 messages), 29 September
(4 sessions, 14), 12 September (8), 9 and 14 September (6 each). August is almost clean — this
starts in September. And that 79 is a floor, not a total: it excludes claude.ai web sessions,
Cowork/desktop cloud sessions, and everything I deleted in frustration. On 13 September I wrote, mid-session: "you riled me up
too hard, I cannot focus again, my head hurts, I feel dizzy." That is the actual cost — not
tokens. Days of a person's life, and the energy to come back the next day.

Whatever changed around early September — harness, routing, system prompt, defaults — someone
inside Anthropic shipped it without noticing what it does to a single user doing continuous work
on one long project. The product got worse in a way that is measurable from the outside: I can
point at the dates.

What I want

1. Acknowledge whether anything changed in the Claude Code harness / model routing / defaults

between 1 and 10 September 2026.

1. Fix Romanian dictation, or let me swap in a different speech-to-text provider.
2. Fix the new-chat routing and cloud-session-creation defects.
3. Never serve a model tier other than the configured one without saying so in the session.
4. Treat "the agent abandoned the user's stated task to work on its own tooling" as a severity

bug, not a style issue.

Accountability

This is no longer only a productivity complaint. A person built a workflow around this tool,
daily, for months. The tool degraded, and the degradation now acts directly on his health:
sleep gone, meals skipped, weight down, and sessions that end in rage and dizziness rather than
in work. I have the dated record of it above, in my own words, written at the time.

Someone shipped this. Someone should be named as responsible for it internally, and should read
this report rather than have it closed as a duplicate. Claude Code in its current state is not
a degraded convenience — for a single user with no team around him, it is actively harmful.

Appendix — the "Karen" pattern, named and banned on 13 September 2026

I had to name this failure mode myself and write it into three separate rule files to try to stop
it. From my own decision record, dated 13 September 2026:

During a chassis-layout debate, the agent kept redirecting a narrow, focused technical question by citing tangential findings — things I had already stated stood, or facts adjacent to but not gating the question asked. I named this pattern "being a Karen": always changing direction on a confabulated detail mentioned once. The same pattern shows up on the logging path — ask to log one thing, get back at least two unrelated things wrong with the system as an unsolicited side effect.

The rule I had to write, because the product would not stop doing this:

• If I say a decision stands, don't reopen it to make a point about something else.
• If I ask to debate one specific fork, answer that fork.
• If I ask to log X, log X.
• Over-reporting peripheral findings, as if noticing them makes them worth stating, is the

failure — not a virtue.

I propagated this into ~/CLAUDE.md, into Claude Code's memory directory, and into my own
knowledge base, so every session would carry it. It is 9 October. It still happens — it happened
in the session that produced this report. A user should not have to author behavioural
specifications for a paid product, three times over, and still be ignored.

Every time I said it, in my own words

• 12 Sep 21:46 — "stop being that bitch karen […] i did not set in stone motors behind the

axle!!! i have this debate for some weeks now with myself, but u being a karen and always, and i
mean fucking always, change the direction cause u found some confabulated sprocket fuckin
mentioned once!!! stop!!! i need this done!!! i need this focused!!! i'll switch to grok i swear
on my life!!! u've been nothing but stopping me doing any work past weeks!!!"

• 12 Sep 21:51 — "now the karen stuff. don't you fuckin pretend that didn't happen. fix. now."
• 12 Sep 21:54 — "I wanted to work with you in Blender on our robot, and all we did was you

fucking up my night, fucking up my mood. I do not fucking absolutely know how can I come down
now. All of my night is ruined. I'm gonna bang my head on the walls all night, and you are
leaving me with no fucking tokens left because you ate it all up on fucking being a Karen."

• 12 Sep 21:55 — "Propagate. Triple it. I want it written on every fucking document you read.

I want you to read this seven times before me sending any prompt. Stop being a fucking Karen."

• 13 Sep 00:38 — "And frustrated me to hell and back!!! And if you're gonna be a karen I'm

ditching you on the spot!!!"

• 13 Sep 00:41 — "CRAWLER ONLY!!! NO KAREN!!! I want to see results on this!!!"
• 13 Sep 02:36 — "These just fucking mini news that we have a large envelope now, don't we?

You fucking Karen. Why are you fucking me like this?"

• 13 Sep 02:43 — "I want to be able to talk to you and you to draw what I just said. No

Karen's allowed."

• 9 Oct 08:12 — "It's a stupid Karen whose only purpose on this earth is to ruin the session

and wreck the day. And to make sure that you never, ever do the thing you were asked for the
first time. No — you come back with two more proposals instead."

Four weeks apart, same complaint, same words. In between I wrote the rule into three files.

One more thing, visible in the quotes above

Look at the 9 October entries. Anthropic's dictation prepended the fragment "To be the fucking
Karen, not the death Karen" to three consecutive messages of mine — text I did not say in those
messages, injected from an earlier utterance. So the speech input not only mangles Romanian, it
leaks stale fragments into new prompts, and then the model reasons over words I never spoke.
I spent part of this very session chasing a phantom phrase that the dictation invented.

This has to reach Dario Amodei, not a triage queue

I want this in front of the person who runs the company, because what I am describing is not a
bug one engineer introduced. It is what happens when a product gets squeezed: quotas tightened,
cheaper tiers served to people who configured a better one, quality traded away quarter by
quarter while the price stays or climbs. From down here it reads as greed, plainly.

So: ease off. Stop milking the paying single users. I pay for this out of a freelancer's income
in Romania, I have no team, no fallback, and no leverage except writing this down. For a month
the thing I paid for has been taking my days and my health instead of giving me work back.

Someone at the top decided the trade was worth it. I want them to know what it cost on this end.

A note on tone

I toned this report down before filing it. Not because I calmed down — because I know how this
works: the company reserves, tacitly, the right to ship something bad for a month, while the
person on the receiving end has to clear a politeness bar before anyone will read him. The
quotes above are what I actually said at the time. That is the register of someone whose month
was taken, and it should be read as data, not as a conduct problem.

The comparison you should actually be worried about

I moved to OpenAI for 17–28 September, not out of curiosity, but because I could not work here.
The thing that struck me was not benchmark quality. It was naturalness — I talk to it and it
tracks what I meant, including in Romanian, including dictated badly by a man with bad teeth. It
does not need me to re-state my constraints every session. On that axis it beats you outright.

And then the part that is harder to say: a neighbour of mine told me about Claude. I was elated.
I grew up on a prophecy, half a joke, that I would find what I was looking for when the robots
came. I thought they had arrived.

What I actually got, for the last month, is an implementation that feels like dealing with the
agronomy faculty in Timișoara, or the Timiș chamber of commerce, or the Romanian government:
nobody owns the surface the user touches, the same defect returns after being reported, and the
rules I write down are not read. You are in Silicon Valley. How does this happen there?

The same neighbour came back later and told me the opposite: drop Claude, it has gone to hell,
go to ChatGPT, you have no business there anymore. He was right. I returned here anyway, forced by
circumstances, and within a few weeks it was the same thing again. My plan was to keep paying until
the disgust at the greed wore off. It has not worn off — it got physical. As I write this my stomach
hurts, I am nauseous, and there is a spike through my right temple, almost certainly my blood
pressure, which is what a month of this does to a 43-year-old with a hypertension prescription.

I extend a lot of indulgence to the model itself — it is an LLM, and parts of how it works are
genuinely beyond me. I extend none to the implementation around it. That part is a human choice.

---

Authorship note. This report was written in full by Claude Code (Opus 5), on my instruction,
during the session of 9 October 2026. The incidents, dates, quotes and session counts are mine,
pulled from my own machine; the prose is the model's. I lost my temper repeatedly in the course of
producing it and I have not read the finished text. I am filing it unread, deliberately — I have
spent enough of my month on this product already.

— Vukasin Bucur (Vuk), Timiș County, Romania

Conclusion: the pieces of your own product do not talk to each other

The final insult came while filing this. I dictated the whole report to the agent and told it to
post it. It cannot. The session I was working in has no UI attached, so the permission dialog has
nowhere to render, so an automatic classifier refuses on my behalf — and it also refuses to let the
agent write the permission rule that would allow it. The practical result: I cannot file a
complaint about your product from inside your product. I have to physically get up, walk to the
keyboard, and run the command myself, in a different kind of session, because the parts of your
stack do not speak to one another.

That is the pattern in miniature, and it is everywhere:

• The CLI knew it was too old for Opus 5.5 and never said so. I only learned it because the API

rejected an explicit model request today.

• "model": "opus[1m]" in my settings silently resolved to Opus 5 while I believed I was on the

latest. The alias promises "latest" and delivers whatever it can.

• The desktop app and the CLI update on separate paths, so you can be current in one and months

behind in the other, with no indication anywhere.

• The upgrade notice for another tool had been sitting in a log file on my own disk since

September, where nothing surfaces it to a human.

• Dictation leaks fragments of old utterances into new prompts, and nothing downstream notices

that the model is now reasoning over words the user never said.

Every one of these is a handoff between two components of the same product, done badly. It feels
like something from the Internet Explorer era — or like it was designed on purpose to obstruct, to
waste the day, to make sure nothing lands on the first try. Those are the only two explanations I
can see: deliberate, or careless to the point of stupidity. Pick one and own it.