Claude Was Broken for 47 Days and Anthropic Said Nothing — Here's the Full Breakdown
What Actually Happened
Claude refused to read a Shrek toy ad. A Hasbro PDF. Totally harmless — except Opus 4.7 threw an AUP violation because the text stream had the words "CHARACTER OR FOR DONKEY UNDERNEATH." That's it. That was the threat. Meanwhile you're paying $200 a month, the tool is supposed to be working, and it's flagging Donkey from Shrek as a content violation.
This guide is the full breakdown. Three bugs, 47 days, one overtuned guardrail system, and what you actually do about it if you're running Claude in your business right now.
---
You Weren't Imagining It
If Claude felt off to you in March and April — slower, shorter, weirdly unhelpful, refusing things it used to handle fine — you weren't crazy. The tool was broken. Not a little degraded. Structurally broken, in three separate ways, for a month and a half.
The frustrating part isn't that bugs happen. Bugs happen at every AI company. The frustrating part is that Anthropic ran these for 47 days at full price with no public acknowledgment. No banner in the dashboard. No email. No status page entry that said "hey, reasoning quality is degraded, we're working on it." Nothing. Developers started comparing notes on Twitter trying to figure out why their Claude Code outputs looked like they came from a different, worse model. Because they did.
LSU's cyber director — paying $200 a month for a Pro subscription — put it plainly: *"basic help with editing tasks will not be rejected."* That's not a power user asking for something exotic. That's someone asking the tool to do the thing it was sold to do. When Gergely Orosz says Anthropic is "burning dev goodwill" and Theo says his last reason to use Claude Code "died in front of him," those aren't drama posts. Those are signals from people who wanted this to work and watched it stop working.
Here's the full breakdown of what went wrong and what it means for how you run AI in your business.
---
The Full Bug Timeline: Three Breaks, One Tool
### 1. Reasoning Effort Got Quietly Downgraded
**What happened:** Sometime around March 4, Anthropic silently cut Claude's reasoning effort from high to medium. This is a real parameter — it controls how much internal "thinking" the model does before it produces a response. High reasoning effort costs more compute. Medium is faster and cheaper on Anthropic's end. The switch happened without any announcement.
**What it looked like:** Your complex prompts started producing outputs that felt shallow. Multi-step coding tasks that used to nail it on the first pass started requiring two or three rounds of back-and-forth. If you were using Claude Code for architecture decisions, refactoring, or anything that needed the model to hold a lot of context at once, you felt this immediately — you just didn't know what caused it.
**Concrete example:** A developer building a SaaS onboarding flow was using Claude Code to map out a database schema that needed to account for multi-tenant permissions. Pre-March 4, Claude would walk through the entire dependency chain unprompted. Post-downgrade, it kept producing schemas that missed edge cases it had previously caught on its own. Three weeks of "is this my prompting or the model?" before the postmortem came out.
**What you gained from knowing this:** If your Claude Code outputs went sideways in March, it wasn't your prompts. Don't rebuild your prompting workflow around a degraded baseline. Go back and re-run the prompts that felt like they stopped working. The model is supposed to be restored.
---
### 2. Cache Clearing Was Wiping the Model's Memory Every Turn
**What happened:** Claude uses a caching system to hold context — the "memory" of what's already happened in a conversation or a coding session. A bug caused this cache to clear on every single turn. Every time Claude responded, it started fresh. No persistent context. No working memory across the session.
**What it looked like:** You'd give Claude a 500-line file and ask it to refactor a function, it would do it fine — then you'd ask a follow-up question and it would respond like it had never seen the file. You'd paste the file again. It would answer. You'd ask another question and it would forget again. Repeat for 47 days.
**Concrete example:** A client of mine was using Claude to maintain a custom GPT-based email triage system — feeding it conversation threads and asking it to flag, categorize, and draft responses. In early March, it started losing track of the categorization rules mid-session. He thought he needed to restructure his system prompt. He rewrote it four different ways. The problem was never the prompt. The cache was just gone.
**What you gained from knowing this:** If you were doing anything in Claude Code that depended on persistent context — long files, multi-step builds, iterative debugging — your frustration was real and it was caused by a real bug, not a real limitation. The architecture you were building toward isn't flawed. The tool was.
---
### 3. A System Prompt Was Capping Responses at 100 Words
**What happened:** A system prompt got deployed — on Claude Code, a coding tool — that instructed the model to keep responses under 100 words. On a coding tool. Where a single function can easily run longer than 100 words just in the code block.
**What it looked like:** You'd ask for a full implementation and get a stub. You'd ask for an explanation and get a sentence. You'd ask Claude to complete a task and it would stop halfway through with no indication that it had truncated the output. It just stopped. You'd say "continue" and it would give you another 100-word chunk and stop again.
**Concrete example:** A marketing agency using Claude Code to generate data pipeline scripts kept getting partial outputs — functions with no closing brackets, loops that cut off before the return statement. Their developers spent two weeks debugging code that was syntactically broken before they realized Claude wasn't finishing its own outputs. They thought it was a context window issue and started breaking tasks into smaller and smaller pieces. It wasn't context. It was a 100-word cap on a coding tool.
**What you gained from knowing this:** If your Claude Code outputs were getting truncated and you built a whole new prompting system around "ask for smaller tasks," you probably overcorrected. The constraint isn't in the model's architecture. Test your original task sizes again.
---
### 4. Opus 4.7's Guardrails Became a False-Positive Machine
**What happened:** Anthropic tuned Opus 4.7's safety guardrails more aggressively — presumably to reduce actual harmful outputs. But the tuning overshot. Hard. False positive refusals went from roughly 2 per month to 30+ in April alone across reported cases. Security researchers got blocked. Legitimate PDFs got flagged. A Hasbro Shrek toy ad threw an AUP violation because a product description contained the phrase "CHARACTER OR FOR DONKEY UNDERNEATH" — which refers to a toy accessory slot, not anything remotely violating.
**What it looked like:** Claude started refusing tasks that were totally routine. Editing tasks. Document summaries. PDF analysis. Anything that touched words that pattern-matched to something on an internal block list, even with zero harmful intent and zero harmful context.
**Concrete example:** A cybersecurity firm using Claude for threat modeling — legitimate, professional, exactly what you'd want AI helping with — started getting refusals on standard security research prompts. Prompts they'd been running for months without issue. The model flagged terms like "exploit path" and "injection vector" even in clearly defensive research contexts. They had to rebuild their entire prompt library to work around guardrails that were supposed to protect against bad actors, not block professionals doing their jobs.
**What you gained from knowing this:** If you're hitting unexpected refusals right now, it's worth testing whether the block is content-based or context-based. Add explicit context about your use case and the professional nature of the request. Don't just rephrase the blocked word — reframe the whole request. And if you're still hitting walls, document it and submit to Anthropic. The false positive rate is a known issue. Your report actually moves the needle.
---
### 5. What This Means for How You Run AI in Your Business
This isn't an "avoid Claude" argument. The postmortem Anthropic dropped is genuinely good — specific, honest, and more transparent than most AI companies get when they break something. Credit for that.
But 47 days is long enough to reshape your assumptions. Here's what to actually change.
**Build a model QA checkpoint into your workflow.** You don't need anything fancy. Once a week, run a consistent test prompt — one you know the expected output for — and compare it against the output you're getting. If something degraded without announcement, you catch it in a week instead of six. This is a 10-minute task. It protects your whole build.
**Don't assume the model is the constraint when outputs degrade.** The instinct when Claude starts underperforming is to fix your prompts. Sometimes that's right. But the three bugs above all produced symptoms that looked like prompt problems. They weren't. Before you rebuild your prompting strategy, check the Anthropic status page and developer forums. Other people will be talking about it if it's a platform issue.
**For guardrail refusals, context is the fix — not censorship.** If Claude is blocking a legitimate request, the solution isn't to obscure what you're asking. Add context: your role, your use case, why this request exists, what the output will be used for. Claude's guardrails are context-sensitive, even when they're overtuned. "Summarize this PDF" performs worse than "I'm a marketing manager reviewing this product catalog for an upcoming campaign. Summarize the key product features in this PDF."
**Document your refusals.** Every time Claude blocks something that shouldn't be blocked, write it down. The model, the date, the prompt context, what got refused. You're building evidence that either tells you your prompt framing needs work or, more usefully, gives you a concrete data set to share with Anthropic or use when evaluating whether this tool still fits your stack.
---
The Real Takeaway
The bugs got fixed. The guardrails are supposedly being recalibrated. But the 47 days happened, and they happened because there was no system in place to catch silent degradation before it compounded into a month and a half of bad outputs at full price.
That's not a Claude problem specifically. That's an AI tooling problem in general. These models update quietly, change behavior without changelogs, and don't come with the kind of version transparency you'd expect from any other piece of software you run your business on. The fix isn't to stop using them. The fix is to stop treating them like static tools and start treating them like something you actually monitor.
You build AI systems to get use. But use only works if the system is actually running. A broken agent doing tasks wrong is worse than no agent at all — because at least with no agent you know the work isn't getting done.
---
Stay Ahead of the Next One
The 47-day Claude bug isn't going to be the last silent degradation in an AI tool you depend on. The next one's already brewing somewhere. The advantage goes to whoever catches it first.
The newsletter (signup below) sends one breakdown per week — what's actually breaking in AI tooling, why the symptoms are misleading, and the specific moves to stay ahead of it. No pitch, no funnel. Just the work.
If you're running AI as core infrastructure in your business and want a second set of eyes on your stack, the application form is the way in. We don't take everyone — we take the builds where there's real signal.