Week 1: LLM Foundations

Core Ideas

• Evolution from n-gram models to transformers
• Foundation models versus task-specific fine-tuned models
• Decoder-only, encoder-decoder, and multimodal architectures
• Latency, cost, and quality tradeoffs in model selection

What You Learn

• Subword tokenization and BPE intuition
• Context window constraints and chunking strategy
• Prompt compression and retrieval-first prompting

# Simple token budget planner
SYSTEM = 400
USER = 1200
RETRIEVAL = 2200
RESPONSE_BUDGET = 1000
TOTAL = SYSTEM + USER + RETRIEVAL + RESPONSE_BUDGET
print(f"Token budget: {TOTAL}")

Patterns

• Instruction-template separation
• Few-shot prompting with compact examples
• Role prompting for domain behavior
• Self-critique and refinement loops

Safety and Reliability

• Schema-constrained generation and validation retries
• Prompt injection and policy checks
• Toxicity and data-leak prevention patterns

Hands-On Labs

• Lab 1: Compare two LLMs for summarization quality
• Lab 2: Build reusable prompt templates
• Lab 3: Add JSON schema output and parser
• Lab 4: Implement token budget checks
• Lab 5: Build a PromptOps assistant

What each concept means in simple language

• Token: A small piece of text the model reads. Think of tokens as tiny text blocks, not full words.
• Context window: The model's short-term memory. If your input is longer than this limit, some details get dropped.
• Prompt: Your instruction to the model. Better instructions give better output.
• Temperature: Controls creativity. Lower means safer and consistent, higher means more creative and variable.
• Guardrails: Rules and checks that keep model answers safe, structured, and reliable.

Exact daily execution plan (2 to 3 hours/day)

1. Day 1: Read the model family overview, then create one-page notes comparing 3 models on cost, speed, and quality.
2. Day 2: Run token counting on 10 prompts. Rewrite long prompts into concise prompts while preserving meaning.
3. Day 3: Write 5 prompt templates: summarization, extraction, rewrite, classification, and QA.
4. Day 4: Add schema validation. Force model output into JSON and retry when output is invalid.
5. Day 5: Build and demo PromptOps assistant that takes raw text and returns summary + key points + action items.

Assignment to complete Week 1

Build a small web or CLI app called PromptOps Assistant. Input: long meeting notes. Output: structured JSON with fields summary, decisions, risks, next_steps. Add one validation rule: if JSON is invalid, auto-retry with correction prompt.

Submission checklist

• One readme that explains your model choice and token strategy
• At least 10 tested prompts with before/after improvements
• JSON validator with retry logic
• Three real test inputs with clean outputs
• Short demo recording (3 to 5 minutes)

How you know you passed Week 1