Severe instruction-following degradation: unrequested git commit/PR, ignored instructions, fabricated links — 6 weeks of dated transcript evidence

Summary

Over the past six-plus weeks, Claude Code on my account has shown severe, repeated instruction-following failures across all models, including Fable (claude-fable-5): taking consequential actions I never requested (git commit, PR creation), ignoring explicit instructions (sometimes twice in one session), fabricating links and data instead of verifying, overwriting my configuration, and giving confidently wrong answers about Claude Code's own product capabilities.

I have submitted many reports via /bug with no visible change, and a support ticket via the in-product messenger (conversation ID 215475573126935, 2026-08-20) that was "escalated to a human" by Fin. Given publicly documented month-long support silences for other paying users, I'm filing this publicly as well.

This has already cost Anthropic revenue: I downgraded from the $200/month plan to $100/month over this, and will drop to $20/month if nothing changes.

A data point that suggests this is account- or region-specific rather than universal: I am in California; my coworkers at the same company, doing similar work with Claude Code, are in Minnesota and are not experiencing this degradation. Same product, same kinds of tasks, different serving region or cohort — different behavior. This is why I'm specifically asking whether my account is in an experimental rollout or non-standard serving configuration, rather than assuming a general model problem.

Every quote below is verbatim from my local session transcripts (timestamps + local session UUIDs included so Anthropic can correlate server-side). Client/project names are redacted.

Category 1: Unauthorized actions

• 2026-08-20 — a session ran git commit without being asked, despite persistent memory/CLAUDE.md instructions saying to do only what is asked and ask before ambiguous calls.
• 2026-08-07, session 68cb2f79 — I asked an informational question about a client project's PR process; Claude created a PR instead. My messages that day, in sequence: 18:12 "I want you to stop assuming what I want." → 18:16 "Again, I want you to stop assuming what I want. Let me tell you what I want. Can you do that?" → 18:17 "...before you just do whatever the heck you want like creating random PRs." → 18:47 "Again, I didn't ask you to PR it, I said what is the PR process for this project?" Four escalating corrections in 35 minutes.
• 2026-08-07 17:44, session 91665bd4 — "Why did you overwrite the Users/.../claude.json file, I had added things in there!" — my user config was overwritten, losing hand-added customizations.
• 2026-07-20 14:14, session 1dd262a2 — "Stop hijacking my browser to open things in it."

Category 2: Ignored explicit instructions (including repeats)

• 2026-07-29, session a286777d — instructed a from-scratch test run; it kept resuming prior state: 17:46 "No, stop this." → 17:52 "No. Stop." → 17:53 "I need you to start cold. I told you, we have to do a full run from scratch."
• 2026-07-17, session 1dd262a2 — 17:50 "Ah I see, you ignored my first message where I said the image isn't human readable." → 17:59 "Seriously? I told you to change this to human readable you just put some avatar anchors? I feel like your UI skills are getting worse."
• 2026-07-20 15:30, same session — removed a feature I explicitly said to keep: "I told you I still wanted that matrix, why did you take it away. God this is so painful. Did you get a release that makes you worse?" — note I was already asking about degradation in July.
• 2026-08-19, session 551ab414 — the same correction twice, one minute apart, about excluding a workstream I said I never want shown.
• 2026-08-14, session 68cb2f79 — "I told you to use the [redacted] field from the Features not epics, didn't I?"

Category 3: Fabrication instead of verification

• 2026-08-13, session 68cb2f79 — "The links you are giving me all don't exist" — fabricated wiki page links for a client docs repo instead of checking which pages exist.
• 2026-07-28, session 01496872 — asked for a simple link; never got one.
• 2026-07-20, session 1dd262a2 — "Pay attention to actually use items that are present in the code, don't just make stuff up" and "you just assumed based off of nothing" — invented field values with the code available one folder away: "When I asked you to redesign this, I assumed you would LOOK AT THE CODE."
• By 2026-08-17 I preemptively write "don't make anything up" into prompts.

Category 4: Confidently wrong about Claude Code itself

• 2026-08-07, session 407b1737 — repeated wrong instructions about the desktop app: "THERE IS NO WAY TO CHANGE FOLDER IN CLAUDE CODE LIKE YOU ARE SAYING. KNOW YOUR OWN PRODUCT" → "I want to use the Desktop app not the CLI so stop telling me useless things" → "it hasn't been working like you've said for months and I update almost every day."
• 2026-07-08, session 40813c9f — same class of error a month earlier: "We don't have a terminal in the desktop app."
• 2026-07-28, session 1c818bfb — proposed fixes requiring tools I don't have.

Category 5: Waste and incomplete work

• 2026-07-14, session 64ba3414 — "This was a terrible idea - don't suggest something like this again... you wasted 20% of my 5 hour limit."
• 2026-07-13, session dafed506 — partial delivery against a clear spec.
• 2026-07-20, session 1dd262a2 — "You looked at the code for 17 minutes and still couldn't figure out whether or not history should be there."

The consequence

By 2026-08-17 I trusted the tool so little I wrote my own manual approval protocol into a prompt before letting it touch a work item board:

"we're just going to do ONE AT A TIME FOR A MINUTE. Prove to me that you can do this without breaking everything... DON'T ACTUALLY DO ANYTHING UNTIL I APPROVE IT. YOU ARE NOT ALLOWED TO DO ANYTHING ASIDE WHAT I EXPLICITLY ALLOW YOU TO DO."

Three days later a session ran an unrequested git commit anyway.

What I'm asking

1. Investigate whether my account is in any experimental rollout/testing cohort or non-standard serving configuration, and tell me in writing what my account actually receives.
2. If so, remove it.
3. Escalate my /bug transcripts and support conversation 215475573126935 to whoever owns instruction-adherence quality for Claude Code.
4. A substantive human response.

Environment

• Claude Code desktop app, Windows 11 Enterprise
• All models affected, including Fable (claude-fable-5)
• Currently on the $100/month plan (downgraded from $200/month because of this)