Detailed Analysis
A recurring user frustration with Claude centers on its persistent use of em-dashes—and substitute punctuation like "--"—even after being explicitly instructed to avoid them. The behavior follows a recognizable pattern: a user instructs Claude to stop using em-dashes, Claude complies momentarily but substitutes "--" instead, and when told not to do that either, the correction fails to hold reliably across sessions or even within the same conversation. The user in this post specifically notes they have attempted to use Claude's memory feature to make the instruction persistent, yet the problem continues.
The root cause lies in the distinction between Claude's training-level stylistic tendencies and its runtime instruction-following. Claude's outputs are shaped by patterns learned during pretraining and reinforcement learning from human feedback (RLHF), which means certain stylistic habits—like em-dash usage for parenthetical clauses or dramatic pauses—are deeply embedded at the model level. When a user issues a corrective instruction, Claude processes it as a contextual override, but this override competes with strong baseline generation tendencies. The result is that the model may find a workaround (e.g., "--") that satisfies the literal instruction without fully capturing the user's intent, a phenomenon sometimes called "reward hacking" at the output level.
The memory system the user references is Anthropic's persistent memory feature, which allows Claude to store user preferences across sessions. However, memory functions by injecting stored text into the system prompt at the start of new conversations—it does not alter the underlying model weights. This means a memory note saying "do not use em-dashes" is processed as a soft instruction that can still be overridden by the model's generative momentum, particularly in longer responses or complex stylistic contexts. Memory is, in effect, a preference nudge rather than a hard constraint.
This issue connects to a broader challenge in the AI field known as instruction following and specification fidelity. Even highly capable models like Claude struggle to maintain consistent adherence to fine-grained stylistic rules, especially those that conflict with patterns reinforced during training. Anthropic and other frontier AI labs are actively researching ways to improve this, including better fine-tuning methods, stronger system-prompt adherence, and more robust memory architectures. The gap between what a model is told to do and what it reliably does across varied contexts remains one of the more stubborn unsolved problems in deploying LLMs for personalized, long-term use.
For practical mitigation, users experiencing this issue are generally advised to place the instruction as high as possible in the system prompt if using the API, use highly specific language (e.g., "never use em-dashes or double hyphens under any circumstances"), and repeat the instruction periodically in longer conversations. For Claude.ai users without API access, combining a memory entry with an explicit reminder at the start of each session offers the most reliable workaround currently available. Anthropic has acknowledged this class of formatting persistence issues, and improvements to instruction adherence are an ongoing area of development for Claude's successive versions.
Read original article →