Spent the last few months wiring up automations across Claude, GPT, Gemini and Kimi through n8n, and the #1 thing that separates a workflow that actually runs unattended from one that breaks every other day isn't the model — it's the prompt structure.

Most people write prompts like they're chatting. For automation, you need prompts that behave like APIs: fixed input schema, fixed output schema, zero ambiguity. A few things that made the biggest difference for me:

1. Always force structured output (JSON schema or XML tags) even if you don't need it yet — future-you will need to parse it eventually.

2. Give the model an explicit "if you're unsure" instruction. Without this, models silently guess instead of flagging low-confidence outputs, which is where automations quietly break.

3. Version your prompts like code. The second you tweak a prompt in production without tracking the change, you lose the ability to debug regressions.

4. Claude Skills > generic system prompts for repeatable multi-step tasks — packaging instructions + context + examples into one reusable unit cuts iteration time dramatically.

If you're deep in the n8n + LLM automation world, curious what's broken on you the most — model drift, rate limits, or just brittle prompts?