Prompt Engineering Mastery Path¶

Mastery Path

BOOTCAMP

VINOD BHANUJA

A prompt is a program written in plain language. Like any program it has inputs, assumptions, edge cases and bugs — and like any program it can be designed carefully, tested against examples, measured, and maintained. Most "prompt tips" you will find online are folklore: magic phrases that seemed to work once, for one person, on one model version. This course is about the opposite: the craft and discipline of prompting. You will learn why a prompt works, how to tell whether a change actually made it better, and how to keep it working when the model underneath changes.

The course is written for two audiences at once. Everyday users — writers, analysts, students, managers — who type prompts into a chat window and want reliable, useful answers. Developers who put prompts inside applications, where a prompt runs thousands of times on inputs nobody has seen yet. Levels 1 and 2 serve both groups; Levels 3 and 4 lean toward people who evaluate, secure and ship prompts, but every lesson explains its ideas in plain terms first.

Level 1 builds the foundations: how a model turns your text into tokens and predicts a continuation, what the parts of a good prompt are, clarity, context, personas and why they are weaker than people think, few-shot examples, formatting, iteration, and the classic failure modes. Level 2 covers the working techniques: reasoning prompts and when they don't help, chaining, structured JSON output, system prompts, style control, extraction, classification, summarization, templates and transformations. Level 3 is about evidence and safety: test sets, rubrics, LLM-as-judge and its biases, A/B tests, prompt injection from a defender's point of view, guardrails, tool-calling, long contexts, images, and sampling settings — ending with a small eval harness you run yourself. Level 4 treats prompts as production artifacts: versioning, regression tests, cost and latency, agents, domain-specific prompting, ethics, team libraries, and knowing when retrieval or fine-tuning beats a better prompt.

Every lesson has a How It Actually Works section explaining the mechanism behind the advice — why examples steer format so strongly, why instructions buried in the middle of a long context get missed, why a model grading its own kind of output is biased, why "think step by step" changes what the model conditions on.

Vendor-neutral, and honest about outputs

The techniques here apply to any modern instruction-tuned model — hosted or open weights. Examples are written as plain prompts you can paste into any chat interface or API. Models change often, and the same prompt can behave differently across models and even across versions of one model, so this course never shows a model response as if it were a recorded output. Where it describes what a model typically does, it says so in words. The Python examples use a small mock model function so they run without an API key; their printed output is real output from running that code.

How the program is organized¶

Level	Focus	Modules
Level 1 · Entry	Tokens & prediction, prompt anatomy, clarity, context, personas, few-shot, formatting, iteration, failure modes	9 topics + 1 project (a reusable prompt with before/after)
Level 2 · Intermediate	Reasoning prompts, chaining, JSON output, system prompts, style control, extraction & classification, summarization, templates, transformations	9 topics + 1 project (support-ticket triage chain)
Level 3 · Advanced	Test sets & rubrics, LLM-as-judge, A/B tests, prompt injection defense, guardrails, tool calling, long context, multimodal, sampling	9 topics + 1 project (a small eval harness)
Level 4 · Master	Versioning, regression tests, cost & latency, agents, code/data/writing prompts, ethics, team libraries, RAG vs fine-tuning	9 topics + 1 capstone (a production prompt package)

How to use this site¶

Keep a prompt notebook. A plain text file per task, with each version of the prompt, what you changed, and what happened. From Level 1 lesson 08 onward the lessons assume you have one.
Try every prompt on a model you have access to. Any capable chat assistant works. If you can, try the same prompt on two different models — the differences teach more than either result alone.
Run the Python. Code examples need only Python 3.9+ and the standard library. Swap the mock model for a real API call when you are ready.
Do the Exercise at the end of each lesson; the module-10 projects build on them.
Related courses: for building LLM applications (API calls, streaming, agents, serving) see LLM Dev Mastery Path; for retrieval pipelines see RAG Mastery Path; for choosing and adopting AI tools see AI Tools Mastery Path; and for one specific assistant in depth see Claude Training Mastery Path. This course deliberately stays on the prompt itself and links to those where they go deeper.