Claude Opus is failing my complex engineering workflows lately

I’m seeing a massive spike in logic failures when using Claude Opus for heavy lifting, and it feels like the model's reasoning capabilities have taken a serious hit. Tasks that used to resolve in a single, clean pass are now turning into endless loops of "fix this, now fix that" because the model is hallucinating context or simply ignoring explicit constraints.

The most frustrating part isn't just the occasional error; it's the loss of instruction following in professional coding workflows. For anyone using a structured setup, this is a nightmare.

The breakdown in coding reliability

When I'm working on a larger codebase, I rely heavily on project-specific rules to keep the LLM agent on track. Recently, Opus has started behaving erratically in several specific ways:

• Ignoring CLAUDE.md rules: I have strict instructions in my project documentation regarding architectural patterns and naming conventions. Opus is now routinely bypassing these, treating my project rules as optional suggestions rather than mandatory constraints.
• Unsolicited file edits: Instead of focusing on the specific function I asked it to refactor, it frequently decides to "clean up" or rewrite unrelated files in the same directory, often breaking existing logic in the process.
• Stale context reasoning: It has started arguing for specific implementations based on outdated comments or old code snippets, even when the current state of the file clearly contradicts them.
• Guesswork over verification: Rather than actually reading the imported modules or verifying the type definitions, it just guesses what the API looks like. This leads to a "blind edit" cycle where it introduces subtle bugs that are hard to spot during a quick review.

The "AI Shrinkflation" problem

It feels like we are witnessing a form of technical shrinkflation. We are paying the same premium subscription fees, but the quality of the output is degrading. This forces a much more expensive workflow: either you burn through your Opus token limit on constant retries to get a single working solution, or you are forced to pivot to even more expensive models or specialized tools just to maintain the baseline quality you used to get from Opus for "free."

There is a suspicion in the dev community that Anthropic might be implementing stricter safety filters that trigger a silent fallback to smaller, less capable models, or perhaps they are throttling compute during peak loads to manage costs. Whatever the underlying cause, the practical result is that the model is becoming less "intelligent" for high-level engineering tasks.

My current workaround

I've actually found myself reaching for other tools for tasks I previously would have handed off to Opus without a second thought. If you are hitting this wall, I'd suggest a few tactical shifts in your AI workflow:

1. Aggressive Context Injection: Since the model is ignoring files like CLAUDE.md, stop relying on the agent to "find" the rules. Manually paste your core constraints into the system prompt or the start of the chat.
2. Smaller, Modular Prompts: Instead of asking for a feature implementation, break it down into tiny, verifiable steps. If the model tries to edit unrelated files, shut it down immediately to prevent a cascade of errors.
3. Model Switching: I've started using a hybrid approach, using more specialized models for pure logic and saving Opus only for high-level architectural brainstorming, though even that is becoming hit-or-miss.

If you've successfully migrated your deep-dive engineering tasks to a different LLM or a different setup, I'm curious to know what your stack looks like now.