Artificial Intelligence & Data Science · 1.5

Prompt Engineer

Master the art and science of designing effective prompts for LLMs to solve complex tasks reliably and safely.

3–5 months4 phases22 skills

Foundations

Understanding LLMshow transformers work (tokenization into subwords, self-attention mechanism, context window as the maximum input+output token count), temperature (0.0 = deterministic for factual tasks, 0.7 = balanced, 1.0+ = creative), top-p (nucleus sampling — consider smallest token set with cumulative probability ≥ p), frequency and presence penalties for reducing repetition, token-based API pricing (input vs output tokens)

Prompt basics

Prompt basicszero-shot (instruction without examples), few-shot (2–5 input-output examples before the task — dramatically improves complex/ambiguous tasks), system prompts (define persona, rules, output format), user prompts (the task), assistant prefilling (pre-fill response start to steer format)

Major LLM platforms and APIs

Major LLM platforms and APIsAnthropic Claude (claude-sonnet-4-20250514 via api.anthropic.com — 200K context, strong instruction following, excellent structured output), OpenAI GPT-4o (via api.openai.com — function calling, JSON mode, vision), Google Gemini 1.5 Pro (1M token context window, multi-modal), open-source models with Ollama for local inference (Llama 3 8B/70B, Mistral 7B, Mixtral 8x7B)

API usage

API usageauthentication with environment variables (never hardcode keys), rate limiting with exponential backoff, streaming responses token-by-token for user-facing apps, cost tracking by logging input/output token counts, error handling (AuthenticationError, RateLimitError, APITimeoutError)

Phase 02 · 8h

Core Prompting Techniques

Chain-of-thought (CoT) prompting

Chain-of-thought (CoT) promptingappend "Let's think step by step" for math/logic/reasoning tasks, provide structured reasoning templates, few-shot CoT (examples that include the reasoning process), self-consistency (generate multiple CoT solutions at temperature=0.7 and take majority vote), when NOT to use CoT (simple lookups, creative writing)

Role prompting and persona design

Role prompting and persona designdefine expertise, communication style, constraints, and output format in the system prompt — test that the persona holds under adversarial inputs

Structured output

Structured outputJSON output with schema specification and Pydantic validation on the receiving end, XML tags for structuring input and output (<context>, <instructions>, <output>), Markdown tables with exact column specifications, schema enforcement using OpenAI's JSON mode (response_format={"type":"json_object"}) and instructor library for Pydantic-validated responses

Prompt chaining

Prompt chainingdecompose complex tasks into sequential API calls (Step 1 output becomes Step 2 input), branching chains (use Step 1 output to select which Step 2 prompt to run), map-reduce for long documents (split into chunks, process each with same prompt, synthesize with a combine prompt), validation chains (second prompt checks first output for errors)

Tree-of-thought and ReAct

Tree-of-thought and ReActToT (generate 3 approaches, evaluate each, proceed with best), ReAct (interleave reasoning with tool actions and observations — foundation of agentic AI)

Prompt templates

Prompt templatesbuild reusable templates with Jinja2 (Template("Analyze this {{ doc_type }}:\n{{ text }}").render()), maintain a versioned prompt library organized by task category, A/B test prompt versions by tracking quality scores

Phase 03 · 14h

Advanced Techniques

RAG prompt design

RAG prompt designcontext injection with source attribution tags (<document source='report.pdf' page='15'>chunk_text</document>), citation instructions ("Cite sources in brackets [filename, page]"), handling retrieval noise ("Some documents may not be relevant — use only relevant information"), extractive vs abstractive QA vs multi-document synthesis patterns

Multi-modal prompting

Multi-modal promptingimage + text (describe what to analyze and specify output format), document understanding (extract structured data from scanned documents/invoices as JSON), chart/graph analysis (describe trends, extract values, identify patterns)

Tool use and function calling

Tool use and function callingdesigning tool schemas with clear names, descriptions (tell the model WHEN to use each tool), and typed parameters, multi-tool orchestration (model calls tools in sequence), error handling for tool failures

Evaluation and benchmarking

Evaluation and benchmarkingbuilding test sets of 50–200 input-output pairs with gold-standard answers, automated metrics (exact match for classification, semantic similarity via embeddings, LLM-as-judge scoring on 1–5 scale with justification), human evaluation with rubrics and inter-rater agreement (Cohen's kappa), red-teaming (adversarial inputs, prompt injection attempts, edge cases)

Prompt optimization

Prompt optimizationDSPy framework (define task as a module, DSPy automatically generates and optimizes prompts), prompt compression (reduce length while maintaining quality), cost-quality tradeoff analysis (cheaper model with elaborate prompts vs expensive model with simple prompts)

Phase 04 · 19h

Domain Applications & Capstone

Code generation prompts

Code generation promptsprovide specifications with language, error handling requirements, and test cases in the prompt; debugging prompts (include error message, code, expected behavior); code review prompts (check for bugs, performance, security, readability); refactoring and test generation prompts

Content creation

Content creationcopywriting with brand voice, target audience, and anti-cliché constraints; style transfer; summarization with exact bullet count and focus areas; translation with cultural adaptation

Data extraction

Data extractionentity extraction returning structured JSON (persons, organizations, dates, amounts), classification with confidence levels, structured data extraction from unstructured text (resumes, invoices, contracts)

Safety and guardrails

Safety and guardrailsinput classification before processing (check for prompt injection, off-topic requests, toxic content), output validation (PII leakage detection with regex, hallucinated URL checking, JSON schema validation), prompt injection defense (delimiters, explicit instructions to ignore overrides), content moderation with provider moderation endpoints, rate limiting and per-user token budgets, logging all prompts/responses for monitoring

Capstone Project 1

Capstone Project 1Build a multi-agent research assistant with LangGraph — web researcher agent + data analyst agent + report writer agent, deployed as a Streamlit application

Capstone Project 2

Capstone Project 2Build a comprehensive evaluation framework for a customer support chatbot — 100-query test set, automated metrics, LLM-as-judge scoring, red-teaming results, prompt optimization report showing quality improvement over 5 iterations

Capstone Project 3

Capstone Project 3Domain-specific prompt library (legal, healthcare, finance, or education) — 15+ versioned templates with documentation, evaluation results, safety guardrails, packaged as a pip-installable Python library

Ready to start Prompt Engineer?

Start this roadmap to track progress and generate courses for each skill. Login required only to begin.