Prompt engineering, "Learn to Spell"
What you’ll learn
Section titled “What you’ll learn”Prompt engineering is the single highest-leverage application skill and the cheapest fix when something goes wrong. The source curriculum is the Full Stack Deep Learning LLM Bootcamp (Spring 2023), by Charles Frye, Sergey Karayev, and Josh Tobin, freely available at fullstackdeeplearning.com/llm-bootcamp with recorded lectures on the Full Stack Deep Learning YouTube channel.
You will apply the prompt-engineering toolkit (clarity and specificity, format constraints, few-shot examples, chain-of-thought, the system prompt as the persistent spec, persona/tone, delimiters between instructions and input, end-placement of critical instructions, negative constraints used sparingly); triage failures into wrong-input (code fix), wrong-output given correct input (prompt fix; the largest and cheapest category), or capability ceiling (retrieval/tools/fine-tuning/different model); version prompts in source control with an explicit prompt_version and test on 20-50 real held-out examples whenever you change them; recognize when prompts run out (missing knowledge to retrieval, missing external systems to tool use, persistent failures cheap to train in to fine-tuning); and use prompt engineering as the first deliberate move against the three productive limits from lesson 2.
Where this fits
Section titled “Where this fits”This is lesson 3 of 11, the last lesson of Phase 1 (foundations and the first app). It closes the smallest complete loop a production builder needs: ship a minimum app (L1), hold a working picture of the constraints (L2), and write the spec the application runs against (L3). Phase 2 opens with augmented language models, the first time the prompt gets context fetched from outside the model.
Before you start
Section titled “Before you start”Prerequisites: lesson 2 of this track (the three productive limits and the system-prompt vs user-message split this lesson builds on). Familiarity with one of the major provider APIs is helpful; the lesson uses provider-agnostic patterns with provider-specific touches noted (e.g. Anthropic’s system parameter is the canonical worked example).
About the math
Section titled “About the math”None. This is a craft and discipline lesson: precise instruction writing, format constraints, few-shot pattern matching, and the engineering practice of versioning and testing prompts. No formulas to derive.
By the end, you’ll be able to
Section titled “By the end, you’ll be able to”The single capability this lesson builds: write effective prompts for production, and recognize when a prompt fix beats a code fix. Concretely, you will be able to:
- Apply the prompt-engineering toolkit (clarity, format, few-shot, chain-of-thought, system prompts, persona, delimiters)
- Triage failures into wrong-input (code fix), wrong-output (prompt fix), or capability ceiling
- Version prompts in source control and test on 20-50 real held-out examples
- Recognize when prompts run out (retrieval, tool use, fine-tuning)
- Use prompt engineering as the first deliberate move against the three productive limits
Time and difficulty
Section titled “Time and difficulty”- Read time: about 13 minutes
- Practice time: about 12 minutes (rewrite a vague prompt with the toolkit, walk a prompt-fix-vs-code-fix triage, plus flashcards)
- Difficulty: standard (no math; the work is internalizing the toolkit and the engineering discipline)