Level 0: AI Foundations
Prerequisites: None Goal: Understand how to engineer systems around LLMs — not just use them
Why This Level Exists
Most AI engineering failures are not LLM failures. They are system design failures — engineers who understood what the model could do but not how to build reliably around it.
This level builds the mental models required before writing a single line of AI code. Skip it and you will spend months debugging problems that have known solutions.
What You'll Learn
- How LLMs process information (and why it matters for engineering)
- The Harness Principle — the deterministic wrapper that makes AI reliable
- The 50/50 Rule — the most important rule in AI engineering
- LLM failure mode taxonomy — know your enemy before you fight it
- Provider landscape — how to choose without locking in
- Cost mental models — reasoning about token budgets as engineering constraints
- Latency tradeoffs — why speed is a first-class design concern
Contents
| File | What It Covers |
|---|---|
| 01-how-llms-think.md | Tokens, attention, context windows as engineering primitives |
| 02-the-harness-principle.md | The deterministic wrapper standard |
| 03-50-50-rule.md | 50% AI reasoning, 50% deterministic code |
| 04-failure-modes.md | Taxonomy of LLM failure modes |
| 05-provider-landscape.md | OpenAI, Anthropic, Google, Azure — decision framework |
| 06-cost-mental-models.md | Token budget thinking |
| 07-latency-tradeoffs.md | Why speed matters and how to reason about it |
| checklists/foundations-readiness.md | Readiness gate before Level 1 |
Core Mental Model
The LLM (purple) is one box. Your engineering (blue) is four boxes. This is intentional.
The Foundational Question
Before building any AI feature, answer these four questions:
- What is the task boundary? — Where does AI reasoning end and deterministic code begin?
- What is the failure cost? — What happens when the model is wrong? Embarrassment? Data loss? Financial loss?
- What is the trust boundary? — What data, tools, and systems can the model access? What can it not?
- What is the evaluation strategy? — How will you know it's working? How will you know when it stops working?
If you cannot answer all four, stop. You are not ready to build.
Anti-Patterns at This Level
❌ "The model is smart enough to figure it out"
Root cause: Conflating intelligence with reliability. LLMs are impressively capable but probabilistically unreliable. Without a harness, every invocation is a roll of the dice.
❌ "We'll add guardrails later"
Root cause: Treating safety and reliability as features rather than architecture. By the time you add them later, you're refactoring your entire system.
❌ "We just need a better prompt"
Root cause: Treating prompt engineering as the primary reliability lever. Prompts are soft constraints. Deterministic code is a hard constraint. Use both.
❌ "We'll evaluate it manually"
Root cause: Not treating AI behavior as a measurable system property. Manual evaluation doesn't scale and doesn't catch regressions.
Enterprise Considerations
At enterprise scale, every item in this level becomes a compliance concern:
- Failure modes map to regulatory risk categories
- Cost mental models become CFO-reported line items
- Provider landscape decisions trigger procurement, legal, and security reviews
- Latency tradeoffs appear in SLA negotiations
Do not skip this level because it feels too basic. Revisit it every 6 months — your understanding will compound.
Readiness Gate
Before proceeding to Level 1, complete the Foundations Readiness Checklist.
Minimum bar: You can explain the Harness Principle, the 50/50 Rule, and three LLM failure modes to a non-technical stakeholder.