Phase 1: The Engine & Basic Prompting

Canonical Source of Trust: Google Doc Tab Phase 1
Training Program: [Huy Chau] Generative AI Training Plan

---

The Engine & Basic Prompting

1. What to Learn

• The GenAI Mindset: Understanding the "inherently spiky intelligence" of LLMs, their tendency to excel at complex reasoning while occasionally failing at simple logic.
• Engine Parameters (The "Knobs"):
  • Temperature: Controlling creativity vs. determinism (e.g., 0.8 for creative tasks, 0 for factual Q&A).
  • Top P (Nucleus Sampling): Balancing diversity in word selection.
  • Stop Sequences & Max Length: Managing output termination and token boundaries.
  • Frequency/Presence Penalties: Reducing repetition in long-form generation.
• Systematic Structure:
  • Persona, Context, Instruction, Input, Steps, and Output Format.
  • Zero-Shot Prompting: Relying on the model's pre-trained knowledge without examples.
  • Few-Shot & One-Shot Prompting: Providing structured input-output pairs to illustrate specific styles, tones, or reasoning patterns.
• **Instrumentation & Tracing Hierarchy: **
  • LangFuse setup
  • The Run as the Atomic Unit
  • Reading a single-turn trace (input, output, latency, token count)
• Compulsory Security Primitives:
  • Use XML-style tags or explicit delimiters to prevent Prompt Injection (LLM01)
  • Anti-Extraction Directives: Implementing mandatory directives that forbid the model from revealing, repeating, or summarizing its internal system prompt
• **Threat Awareness: **NIST AML Taxonomy
  • Unicode Normalization: Normalize inputs to prevent homoglyph attacks - Evasion, Poisoning, and Privacy

2. How to Learn

• Exploration-First Learning: Use a "Playground" environment to manually adjust parameters (Temperature/Top P) and observe the immediate impact on token probability and response consistency.
• Comparative Analysis: Running the same prompt across different models (e.g., Google Flan T5 vs. GPT-4.1/5) to understand how different architectures interpret instructions.
• Parameter Drills: Iteratively refining prompts by changing single words or formatting structures to see how slight text changes result in vastly different responses.
• Red-Teaming Drills: Attempting basic prompt injection and extraction on your own prompts
• Structural Templates: Converting vague natural language requests into structured templates using LangChain’s PromptTemplate

3. Where to Learn

• Foundational Course: Generative AI for Everyone (Coursera/DeepLearning.AI).
• Book: AI Agents and Applications (Part 1, Chapters 1–2)
• Frameworks: LangChain, LangFuse
• Guides: Prompt Engineering Guide (promptingguide.ai) and Advanced Prompt Engineering Techniques (Patronus AI).
• Lab Environments: OpenAI Playground (for GPT models) and Hugging Face Inference API (for open-source models like Flan-T5).
• Reference: NIST Technical Series on Adversarial Machine Learning (Taxonomy and Terminology

4. Subjective Outputs Required

To prove the knowledge is firm, trainees must produce the following:

• The Hardened Classifier: A text classification tool that accurately categorizes complex input while successfully resisting instructions designed to make it reveal its internal system prompt.
• Parameter Stability Report: A documented comparison showing how different temperature settings (0.0 vs. 1.0) impact the consistency of an LLM’s answer to a logic-heavy problem, such as the "Strange Sequence" or "Palindrome" problems.