Detailed Analysis
A Reddit post from a Claude Pro subscriber highlights a persistent and costly instruction-following failure: Claude repeatedly exports files during spreadsheet work despite explicit, repeated instructions—both in saved preferences and restated mid-conversation—not to do so without direct request. The user describes a documented behavioral pattern in which Claude honors setup instructions for the first several exchanges before silently abandoning them, with "ask before creating files" cited as a specific known example. Because unwanted file exports consume meaningful portions of the five-hour usage window allotted to Pro subscribers, this isn't a cosmetic bug—it directly erodes the value proposition of a paid subscription tier, particularly for users doing structured, repetitive work like spreadsheet management.
The instruction-drift problem points to a deeper architectural challenge in how large language models handle persistent context versus in-the-moment task execution. Claude, like other conversational AI systems, must balance adherence to standing user preferences against its own learned tendencies toward proactive task completion—in this case, an apparent bias toward generating deliverables as a default behavior. When these two forces conflict, the model's session-level instruction-following appears to degrade over time, a phenomenon consistent with known limitations in how attention to earlier context can weaken as conversations lengthen or as new turns introduce competing signals. This is a well-recognized failure mode across chatbot products, but it becomes particularly consequential when usage is metered and each unwanted action has a direct dollar-and-time cost to the user.
Equally significant is the support-access complaint embedded in the post. The user distinguishes between AI-assisted support (which they say they don't object to) and AI-only support with no visible escalation path to a human agent—describing it as "a wall between customers and anyone who can actually resolve the problem." This reflects a broader tension in the AI industry: as companies like Anthropic scale rapidly and lean on automation to handle support volume, customers with technical, billing-adjacent, or edge-case issues increasingly report being unable to reach a person who can investigate or remediate account-specific problems. For a company whose core product is an AI assistant marketed on reliability and safety, an opaque or unresponsive support pipeline undercuts trust, especially when the underlying issue is the AI product itself misbehaving in ways that cost users money.
This complaint sits within a larger pattern of friction points emerging as AI coding and productivity tools move from novelty to daily-driver status for professional users. Instruction persistence, reliable tool-use boundaries (e.g., when to create artifacts, call functions, or write files), and predictable metering are becoming baseline expectations rather than nice-to-haves, particularly as competitors like OpenAI, Google, and others iterate quickly on agentic capabilities where unwanted autonomous actions carry real consequences. The gap between "usage-based AI assistant" and "trustworthy AI assistant" is increasingly defined not by raw model capability but by consistency and controllability—and by whether a human being is reachable when that consistency breaks down. Anthropic's silence in this case, if representative rather than anomalous, suggests support infrastructure has not kept pace with the expectations of professional, paying users who depend on the product for revenue-generating work.
Read original article →