To support the development of computational thinking (CT) skills and to promote self-efficacy (SE) in computer science, we present BarrelBots, a puzzle-based block coding activity in Minecraft designed for upper elementary and middle school learners. BarrelBots leverage an appealing and emotionally engaging agent that is controlled by a custom programming language. Learners solve a range of pre-designed puzzles of increasing difficulty. Solving the puzzles requires understanding of basic coding control structures (loops, conditionals, and functions) as well as algorithm design and sequential programming. After mastery of these basic skills, learners are given the opportunity to design new puzzles and share them with peers. To better understand the needs for groups who are historically underrepresented in STEM, we present an initial analysis of data of BarrelBots with middle school participants with low-SES backgrounds (n = 18). More specifically, we present an analysis of three puzzles created by each participant in a summer camp setting. We adapt previously validated scoring schemes to assess creativity and to code feedback given by ChatGPT-4o. Our goal is to investigate the application of AI-based models for scoring and feedback of the artifacts. Our findings suggest that ChatGPT can identify Minecraft block types and structures, but that a zero-shot approach may not be enough to score creative work products similar to humans. The most frequent codes that arose from the LLM feedback show promise for improving the participant-designed puzzles and promoting resilience for our participants, which could have implications for developing CT skills and SE in STEM.