Part III — Beyond the Stack
The stack changes every six months. These seven skills don't.
If you've shipped a few production AI features and you're starting to ask "why am I always re-learning the same lessons with each new framework?" — that's the signal that the leverage has moved from more tools to more intuition.
Junior AI engineers are graded on whether the LLM call returns the right shape. Mid-level AI engineers are graded on whether the system as a whole — prompt, retrieval, evals, observability, fallbacks, cost — is the right design for the problem. The seven skills in this part are how you make that shift.
What's in this part
- Prompting as craft — Context engineering: the entire input as designed artifact.
- Eval mindset — How to think about measurement; LLM-judge biases and mitigations.
- Retrieval quality — Why chunking and hybrid search dominate embedding choice.
- Agent discipline — When agents are the right tool; when they aren't; how to keep loops safe.
- Cost intuition — Order-of-magnitude estimation; why caching is the highest-leverage optimization.
- Latency intuition — TTFT vs total time; streaming UX; perceived vs measured speed.
- Safety mindset — Prompt injection, data exfiltration, supply-chain risk; defense-in-depth.
How to read this part
Slowly. None of these are crash courses — each one is a multi-year skill. The page for each is a curated entry point: the load-bearing ideas, the common failure modes, what to practice, where to read deeper. None of these survive being skimmed.