We used a base prompt template but absolutely had to tweak it for Haiku, and a bit for Sonnet. The syntax was the same, but Haiku responds better to extremely explicit, almost rigid formatting instructions in the system prompt. Think "Return a JSON object with exactly these keys" rather than "Please provide the data in JSON format."
GPT-4 was more forgiving with vague prompts, but you pay for that flexibility. The tweaking time for Haiku was maybe an hour upfront, and it's been solid since.
Spot on about the nuance gap. It's that exact tipping point where you go from parsing to reasoning. I've seen Haiku do brilliantly on extracting error codes and timestamps from a messy stream, but ask it "which of these five warnings is most likely to cause a user-facing delay?" and it starts guessing.
For us, the sweet spot became a hybrid. We use Haiku as a first-pass filter to extract and structure the raw logs, then we have a separate, smaller set of rules (sometimes even just heuristics) that run on the structured output to do the actual prioritization and flag the "need a human" cases. That way you're only paying the big model tax on the 2% of logs that actually require understanding.