breaking down unstructured terminal output into structured HTML via agent skill is a killer pattern. the trick is keeping the skill granular enough that LLM context doesn't explode but complex enough to handle real messy output. are you using vision models for pattern matching on the raw text, or parsing semantically first?