Phase 1: Interactive Flashcard Hub

Active Recall & Spaced Repetition: Master the foundational concepts, parameter mechanics, systematic prompt structures, tracing hierarchy, and security primitives of Phase 1: The Engine & Basic Prompting.
Curriculum Alignment: docs/plan/01_phase1_engine_and_prompting.md
Interaction: Click any card or press Space to flip. Use ← Prev and Next → to navigate.

---

🗂️ Deck 1: The GenAI Mindset

Master autoregressive mechanics, the spiky intelligence paradox, BPE subword tokenization, and fixed compute budgets:

🗂️ Deck 1: The GenAI Mindset

Card 1 of 5

The GenAI MindsetClick or press Space to flip ↺

What is an LLM fundamentally, and what is its mathematical core equation?

P(x_t | x_<t; θ)

Question / Scenario

The GenAI Mindset • AnswerClick to flip back ↻

An LLM is a probabilistic autoregressive next-token predictor. Given a context window of previous tokens x_<t and billions of frozen parameters θ, it calculates a continuous probability distribution over a discrete vocabulary (32k–128k tokens) to sample the next token.

💡 Architect Takeaway: LLMs are statistical sequence generators, not deterministic symbolic knowledge bases.

Technical Explanation

---

🗂️ Deck 2: Engine Parameters (The "Knobs")

Master Temperature scaling, Top-P nucleus sampling, Stop Sequences, Max Length, and Penalties:

🗂️ Deck 2: Engine Parameters (The Knobs)

Card 1 of 7

Engine ParametersClick or press Space to flip ↺

How does Temperature (T) scale logits mathematically, and what is its behavior as T → 0?

P(w_i) = exp(z_i / T) / Σ exp(z_j / T)

Question / Scenario

Engine Parameters • AnswerClick to flip back ↻

Temperature divides raw logits before the softmax function. As T approaches 0 (greedy decoding), the probability of the maximum logit approaches 1.0 while all other tokens drop to 0.0, resulting in deterministic argmax token selection.

💡 Architect Takeaway: Use T=0.0 for structured JSON extraction, code generation, and classification where determinism is critical.

Technical Explanation

---

🗂️ Deck 3: Systematic Structure & In-Context Learning

Master the 6 canonical prompt components (Persona, Context, Instruction, Input, Steps, Output Format), Zero/Few-shot learning, and LangChain templates:

🗂️ Deck 3: Systematic Structure

Card 1 of 6

Systematic StructureClick or press Space to flip ↺

What are the 6 canonical components of a systematic prompt?

Question / Scenario

Systematic Structure • AnswerClick to flip back ↻

1. Persona: Identity, tone, and operational boundaries. 2. Context: Background scenario, domain rules, and axioms. 3. Instruction: The core directive or transformation task. 4. Input: Untrusted user payload enclosed in explicit delimiters. 5. Steps: Procedural execution logic the model must follow. 6. Output Format: Exact schema, JSON keys, or structure required.

💡 Architect Takeaway: Organizing prompts into these 6 components eliminates ambiguity and schema drift.

Technical Explanation

---

🗂️ Deck 4: Instrumentation & Tracing Hierarchy

Master LangFuse tracing, The Run as the atomic unit, single-turn trace inspection (TTFT/ITL), and context propagation:

🗂️ Deck 4: Instrumentation & Tracing Hierarchy

Card 1 of 5

Instrumentation & TracingClick or press Space to flip ↺

Why is semantic tracing necessary when monitoring LLM applications?

Question / Scenario

Instrumentation & Tracing • AnswerClick to flip back ↻

Traditional APM only checks HTTP status codes. An LLM API call can return HTTP 200 OK while outputting a catastrophic hallucination, violating safety guardrails, or failing schema validation. Semantic tracing inspects the actual prompt, completion, and reasoning quality.

💡 Architect Takeaway: HTTP 200 OK does not indicate semantic correctness in generative AI systems.

Technical Explanation

---

🗂️ Deck 5: Compulsory Security Primitives & Threat Awareness

Master OWASP LLM01 injection defenses, XML delimiters, Anti-Extraction directives, Canary tokens, and NIST AML Unicode normalization:

🗂️ Deck 5: Compulsory Security Primitives & Threat Awareness

Card 1 of 7

Security PrimitivesClick or press Space to flip ↺

What is the difference between Direct and Indirect Prompt Injection (OWASP LLM01)?

Question / Scenario

Security Primitives • AnswerClick to flip back ↻

Direct Injection (Jailbreak) occurs when the end user supplies malicious input overriding system instructions. Indirect Injection occurs when untrusted third-party data (web pages, PDFs, emails, database records) read by the model contains hidden adversarial directives.

💡 Architect Takeaway: Indirect injection is the #1 vulnerability in RAG pipelines and autonomous tool-using agents.

Technical Explanation

---

🎯 Active Recall Protocol for Trainees

1. Daily Spaced Repetition: Run through 1 deck daily during Phase 1.
2. Target Mastery: Aim to mark $\ge 80\mathrm{\%}$ of cards in each deck as Mastered before beginning Phase 1 Lab implementations.
3. Cross-Reference:
   • For parameter deep dives: review 02_engine_parameters.md.
   • For full conceptual hierarchy: explore the Phase 1 Architecture Mindmap.