ARM Cortex-M

Starting from a naive vibe coding attempt that exposes how LLMs confuse pin allocations for newer hardware, this post walks through iteratively building a working BME280 weather station on the nRF54L15-DK. We explore how context quality affects code generation and show that even correct results degrade as underspecified details drift between iterations. The takeaway is that both prompt clarity and detail persistence matter as much as model capability.

This opening post introduces the 5-part series and discusses how agentic LLM coding is shifting our workflows to empower domain experts. It examines key challenges: the gap between vague natural-language prompts and precision firmware needs, how rapidly evolving interfaces cause LLMs to generate plausible but incorrect code, and the hard limits where LLMs currently aren’t helpful. The nRF54L15 SPI errata example illustrates how subtle hardware differences can turn confident LLM output into silent data corruption.