In this work, we investigate whether detailed, dataset-specific annotation guidelines can replace or augment annotated few-shot exemplars for prompt-based entity detection. To this end, we propose an in-context learning framework that conditions the LLM on a guideline document, task instructions, and optional demonstrations, enabling single-pass, primarily zero-shot nested entity detection. We evaluate multiple instructional aspects from the comprehensive ACE 2005 annotation guidelines, using three guideline-compliant datasets as benchmarks and GPT-5 variants for inference. Based on extensive experimentation, we demonstrate that detailed, dataset-specific instructions, especially those governing mention boundaries, consistently and robustly improve performance, even in conjunction with established few-shot methods. Indeed, even under zero-shot constraints, our guideline approach outperforms state-of-the-art few-shot methods on ACE-2005, achieving up to 81.2 percent F1 without relying on existing labeled datasets. Our findings present detailed annotation guidelines as a promising resource for improving instruction-based information extraction.