Volume 12 closed with specialization via weights—adapters, LoRA, QLoRA, prefix tuning, and prompt tuning. Those methods change (or prepend) parameters so a frozen backbone behaves like a specialist.
Volume 13 opens the complementary lever: discrete prompting—natural-language instructions that steer the same model at inference time without touching weights. This lecture is the map for Module 13.1. Everything that follows—zero-shot, one-shot, few-shot, chain-of-thought, tree-of-thoughts, self-consistency, reflection, and role prompting—is a craft pattern on top of the basics introduced here.
Learning Objectives
By the end of this lesson, students should be able to:
- Contrast PEFT/adapters (train-time specialization) with prompting (inference-time control).
- Name the core parts of a production prompt: task, context, constraints, format, and success criteria.
- Write a clear, specific instruction that states audience, length, and output shape.
- Separate system-level policy from user-level task wording (preview of later lectures).
- Spot vague prompts and rewrite them into testable specifications.
- Choose when to prompt first versus when to fine-tune or add adapters.
Basic prompting is the practice of steering a language model with carefully written natural-language text—instructions, context, and format rules—so the model produces useful, reliable outputs without changing its trained weights.
Two Layers of Control
After Volume 12 you already know how to specialize a model by training small packs of parameters. Prompting is the application layer sitting above that stack: same API call, different text in the context window.
| Layer | What you change | When it runs | Typical cost |
|---|---|---|---|
| Full FT / SFT | Many or all weights | Train time | High compute & storage |
| PEFT adapters / LoRA | Tiny parameter packs | Train time | Low storage; still needs data |
| Soft prompt / prefix tuning | Learned continuous tokens | Train time | Tiny packs; context tax |
| Discrete prompting | Readable text only | Inference time | Tokens + iteration time |
Rule of thumb: prompt first for format, tone, and light task shaping; reach for PEFT when prompting cannot close a reliable quality gap or when you need a durable specialist shared across many calls.
Anatomy of a Strong Prompt
Task
- What to do (verb + object).
- Who the audience is.
- What “done” looks like.
Context
- Facts the model must use.
- Source text, schema, or policy.
- What to ignore.
Constraints & Format
- Length, tone, banned content.
- Structure (bullets, JSON, table).
- Explicit “if unknown, say so.”
Weak vs Strong Prompt
Vague prompts force the model to invent missing requirements. Specific prompts make failures measurable—essential later in prompt evaluation.
A Practical Template
System vs User (Preview)
Chat APIs often split messages into roles. Put durable policy and persona in the system prompt; put the changing task and documents in the user prompt. Module 13.1 returns to that split in detail after the core techniques.
Task + success criteria
Paste or retrieve facts
Format, tone, limits
Test failures, tighten text
Strengths and Tradeoffs
Strengths
- No training pipeline; ship in minutes.
- Inspectable and editable by non-ML staff.
- Composes with any PEFT-specialized base.
Tradeoffs
- Uses context tokens every call.
- Brittle if left vague or contradictory.
- Cannot teach brand-new knowledge as reliably as SFT on domain data.
“Prompting replaces fine-tuning.” They solve different problems. Prompting controls behavior for this request. Adapters and SFT change default behavior across many requests. Production systems often use both: a lightly adapted model plus carefully engineered prompts, later wrapped with guardrails.
Knowledge Check
- Short Answer: What does Volume 13 add that Volume 12 adapters do not? Answer: Inference-time discrete (text) control without changing weights.
- True/False: Basic prompting requires updating model parameters. Answer: False.
- Multiple Choice: The best first step for a vague product ask is usually: (a) full FT, (b) clarify task/format in a prompt, (c) train from scratch. Answer: (b).
- Short Answer: Name three parts of a strong prompt. Answer: Task, context, and constraints/format (audience/success criteria also acceptable).
- True/False: Soft prompt tuning and typing a natural-language prompt are the same technique. Answer: False—one learns continuous embeddings; the other is discrete text.
- Multiple Choice: Durable persona/policy belongs mainly in: (a) system prompt, (b) random user message only, (c) the tokenizer vocab. Answer: (a).
- Short Answer: Why is “Summarize this” a weak prompt? Answer: It omits audience, length, format, and success criteria.
- True/False: Prompting can be used on top of a LoRA-adapted model. Answer: True.
- Multiple Choice: Prefer PEFT over prompting alone when: (a) you need a durable specialist the prompt cannot reliably achieve, (b) you only need five bullets once, (c) you lack a keyboard. Answer: (a).
- Short Answer: What should the model do when required facts are missing? Answer: State insufficient information instead of guessing (if instructed).
Key Takeaways
- Prompting is application-layer control; adapters/PEFT are train-time specialization.
- Strong prompts specify task, context, constraints, and output format.
- Rewrite vague asks into testable specifications before adding complexity.
- Next: Zero Shot—solving tasks with instructions alone.
Hands-on idea: Give students one messy business email and three rewrite targets (PM bullets, legal risk list, customer reply). Compare outputs from a one-line prompt versus the strong template.
Discussion prompt: When would you ship a LoRA adapter instead of iterating on prompts for another week?
Recap: Basic prompting turns goals into precise, checkable instructions. Continue with Zero Shot.