This paper describes the system submitted by Team Gazoo! for Task 7 of the #SMM4H-HeaRD 2026 shared task on detecting self-reported clinical and social impacts of nonmedical opioid use in social media text. We present a zero-training, prompt-only approach that uses a large language model (GPT-5.4) with structured few-shot prompting and autonomous, iterative rule optimization. Our system encodes a domain-specific entity ontology, three core decision rules, and 65 cognitively organized few-shot examples into a single prompt, with BIO constraint enforcement applied as post-processing. Crucially, the prompt itself is refined by the LLM: at each iteration the model analyzes its own errors and proposes targeted edits to its rules and examples. Through 18 such self-refinement cycles, our system achieved an F1-Strict of 0.53 and F1-Relaxed of 0.60 on the test set, ranking first among all participating teams under both evaluation criteria.