The right level of automation
Not every step in a workflow should be automated.
The useful question isn't "can AI do this?" - it's "what's the cost of AI getting this wrong?"
Low cost of error, high volume, predictable input: automate fully.
High cost of error, low volume,
Founder at Karthea - AI Stack Audits for AI-Native Operators
- System prompts are your most valuable IP If you've built something that works - a workflow, a tone of voice, a classification system, a customer-facing agent - the system prompt behind it is the asset. Treat it like one. Back it up. Version it. Don't let it live only in the
- The model doesn't know what day it is Every AI model has a training cutoff. After that date, it knows nothing new. This matters more than most operators realize. If your AI workflow involves anything time-sensitive - current pricing, recent events, live data, regulatory
- Temperature is a setting, not a mystery Most people who use AI APIs have never changed the temperature setting. Temperature controls how "creative" vs. how "predictable" the model is. Zero means near-deterministic output. One means more varied, less reliable. For
- What "hallucination" actually means in practice Hallucination gets talked about like it's a bug that will eventually be fixed. It's not. It's a fundamental property of how language models work. The practical implication: any AI output that requires factual accuracy needs a

