Prompt engineering: get great results from AI
This curriculum takes a complete beginner from "what is AI?" to confidently designing advanced prompts and real-world AI workflows. Each stage builds on the last: first you develop AI literacy and intuition, then you master practical prompting techniques, and finally you integrate AI deeply into professional and creative work.
AI Literacy & Foundations
New to itUnderstand what modern AI actually is, how large language models work at an intuitive level, and why prompt quality matters — without needing a technical background.
▸ Study plan for this stage
Pace: 4–5 weeks total: Weeks 1–3 for "Co-Intelligence" (~25–30 pages/day, reading in focused 30-minute sessions), Weeks 4–5 for "The Age of A.I." (~20–25 pages/day, slower pace to absorb its denser, philosophical arguments). Plan for one reflection day per week with no new reading.
- What LLMs actually are: Co-Intelligence frames AI as a 'alien mind' — a statistical pattern-matcher trained on human text, not a thinking being, helping readers shed both over-anthropomorphization and dismissiveness
- The 'Jagged Frontier' (Mollick): AI capability is uneven — superhuman in some tasks, surprisingly weak in adjacent ones — so learners must map where AI helps vs. misleads in their own work
- Humans as the essential layer: Mollick's core argument that AI augments rather than replaces when humans stay 'in the loop' as editors, critics, and directors of AI output
- Prompt quality as leverage: Co-Intelligence demonstrates through examples that the specificity, context, and framing of a prompt directly determines output quality — the human's input is the primary variable
- AI as a general-purpose technology (Kissinger): like the printing press or electricity, AI restructures not just tasks but entire ways of knowing, deciding, and governing
- Epistemological disruption (Kissinger): AI systems can reach correct conclusions through processes humans cannot audit or fully understand, challenging Enlightenment assumptions about reason and accountability
- The distinction between narrow AI tools and emergent LLM behavior: both books together illustrate the leap from rule-based automation to systems that generalize across domains
- Responsible use and societal stakes: Kissinger grounds the personal productivity framing of Mollick in larger questions of truth, trust, and institutional integrity
- In your own words, how does Mollick's 'Jagged Frontier' metaphor explain why AI aces a hard task but fails an easy one — and what does that mean for how you should verify AI output?
- Mollick argues you should treat AI like a 'brilliant friend with all the knowledge in the world.' What are the benefits AND the dangers of that mental model, and how does Kissinger's perspective complicate it?
- What does Kissinger mean when he says AI challenges the Enlightenment model of reason? Give a concrete example of a situation where an AI's correct answer arrived through an unexplainable process.
- Based on Co-Intelligence, what are the three to four elements of a prompt that most reliably improve output quality, and why does each element matter?
- Both books were written by non-engineers for general audiences. What does each author identify as the single greatest risk of AI adoption — and do those risks conflict or reinforce each other?
- How would you explain to a skeptical colleague — using ideas from both books — why learning to write better prompts is a meaningful professional skill and not just a tech gimmick?
- 'Jagged Frontier' personal map: After finishing Co-Intelligence, list 10 tasks from your own daily work or study. Use a free AI tool (ChatGPT, Claude, etc.) to attempt each one, then rate AI performance as Strong / Weak / Surprising. Compare your predictions to the results — this makes Mollick's concept visceral and personal.
- Prompt iteration log: Pick one real task (drafting an email, summarizing an article, brainstorming ideas). Write a bare-bones prompt, record the output, then rewrite the prompt adding role, context, format, and constraints. Do this for 3 rounds and write two sentences on what changed — directly practicing the prompt-quality argument at the heart of Co-Intelligence.
- Brilliant friend vs. auditable expert: Choose a factual question in a domain you know well. Ask an AI for an answer, then fact-check every claim. Write a short note on which claims were correct, which were wrong, and which were unverifiable — connecting to Kissinger's point about epistemic accountability.
- Newspaper front page exercise: After finishing The Age of A.I., write a fictional 200-word newspaper headline and lede from the year 2035 that illustrates one of Kissinger's societal concerns. Then write a counter-headline showing a positive outcome. This forces engagement with the book's long-range arguments.
- Vocabulary & concept card deck: As you read both books, create a running glossary of 15–20 terms (e.g., 'large language model,' 'hallucination,' 'jagged frontier,' 'general-purpose technology,' 'epistemic'). Write each definition in your own words without jargon — a reliable test of genuine understanding.
- Synthesis discussion or journal entry: After both books, write or discuss for 15 minutes: 'Mollick says lean in and experiment; Kissinger urges caution and humility. How do I personally reconcile these two stances as I start using AI tools more deliberately?' This integrates both authors' frameworks before moving to hands-on prompt engineering.
Next up: Having built an intuitive mental model of what AI is and why prompts matter (the 'why'), the reader is now primed to move into the practical mechanics of prompt construction — learning the 'how' through structured techniques like few-shot prompting, chain-of-thought, and role assignment.
The perfect first book: a clear-eyed, practical introduction to living and working with AI written by a leading researcher. It builds the mental model you need before touching a single prompt.
Broadens your understanding of what AI means for society and human thinking, giving you the 'why it matters' context that makes everything else more meaningful.
Prompt Engineering Essentials
New to itLearn the core techniques of prompt engineering — instructions, roles, context, examples, and chain-of-thought — and be able to get reliably useful outputs from ChatGPT and Claude.
▸ Study plan for this stage
Pace: 6–8 weeks total. Week 1–3: "The Art of Prompt Engineering with chatGPT" by Nathan Hunter (~20–25 pages/day, including hands-on practice sessions every 2–3 days). Week 4–8: "Prompt Engineering for Generative AI" by James Phoenix (~25–30 pages/day, with dedicated lab days on weekends to experiment wit
- Prompt anatomy: the role of instructions, context, input data, and output indicators as the building blocks of any prompt (Hunter)
- Role/persona prompting: assigning ChatGPT a specific role or identity to shape tone, expertise level, and response style (Hunter)
- Iterative prompt refinement: treating prompt writing as a feedback loop — draft, test, diagnose, and revise — rather than a one-shot task (Hunter)
- Few-shot and zero-shot prompting: using worked examples inside the prompt to steer output format and quality, versus relying on the model's base knowledge alone (Hunter & Phoenix)
- Chain-of-thought (CoT) prompting: instructing the model to reason step-by-step before giving a final answer to improve accuracy on complex tasks (Phoenix)
- Temperature and parameter control: understanding how generation settings affect creativity vs. determinism in outputs (Phoenix)
- Prompt patterns and reusable templates: building a personal library of prompt structures for recurring task types such as summarization, classification, and code generation (Phoenix)
- Model-agnostic principles: recognizing which techniques transfer across ChatGPT, Claude, and other LLMs versus which are model-specific (Phoenix)
- After reading Hunter, can you explain the four components of a well-structured prompt and give an original example of each?
- What is role prompting, and how does assigning a persona change the model's output compared to a plain instruction? Provide a before-and-after prompt pair.
- What is the difference between zero-shot, one-shot, and few-shot prompting, and when should you choose each approach according to Phoenix?
- How does chain-of-thought prompting work, and for what types of tasks does Phoenix recommend it most strongly?
- How would you diagnose and fix a prompt that keeps producing vague or off-topic responses? Walk through the iterative refinement process described by Hunter.
- Which prompt engineering techniques covered in Phoenix are model-agnostic, and how would you adapt a ChatGPT prompt for use with Claude?
- Prompt dissection log (Hunter): For the first two weeks, take 5 real prompts you use daily and rewrite each one explicitly labeling the instruction, context, role, and output format components. Compare outputs before and after.
- Role-prompting A/B test (Hunter): Pick one task (e.g., explain a technical concept). Write three versions of the prompt — no role, a generic role ('you are an expert'), and a highly specific role ('you are a senior data scientist explaining to a non-technical manager'). Document the differences in tone, depth, and usefulness.
- Few-shot template builder (Hunter & Phoenix): Choose three recurring tasks (e.g., email drafting, summarization, brainstorming). For each, craft a few-shot prompt with 2–3 examples embedded, then save them as reusable templates in a personal prompt library.
- Chain-of-thought challenge (Phoenix): Take five multi-step problems (math word problems, logical puzzles, or planning tasks). Run each twice — once with a direct-answer prompt and once with 'Let's think step by step.' Record where CoT improved accuracy and where it made no difference.
- Cross-model porting exercise (Phoenix): Take your three best prompts from the Hunter exercises and run them verbatim on both ChatGPT and Claude. Note where outputs diverge, then adjust wording to make each prompt perform well on both platforms.
- Personal prompt pattern library (Phoenix): By the end of the stage, compile a structured document of at least 10 tested, annotated prompt templates covering different task types (summarization, classification, creative writing, Q&A, code). Include the pattern name, use case, the template itself, and one real output example.
Next up: Mastering these foundational techniques — structured prompts, role assignment, few-shot examples, and chain-of-thought — gives you the reliable baseline output quality needed to tackle more advanced topics such as prompt chaining, autonomous agents, and integrating AI into real workflows in the next stage.
A hands-on, beginner-friendly guide that walks through the fundamental prompting patterns step by step, giving you a repeatable toolkit to apply immediately.
Goes deeper into structured prompting strategies — few-shot examples, chain-of-thought, and system prompts — building directly on the basics you just learned.
AI for Real Work & Everyday Life
Some backgroundApply AI tools to concrete tasks — writing, research, decision-making, and productivity — and develop personal workflows that save real time and effort.
▸ Study plan for this stage
Pace: 8–10 weeks total: Week 1–2 — "Superhuman Innovation" (~25–30 pages/day); Week 3–5 — "Power and Prediction" (~20–25 pages/day, slower pace for denser economic frameworks); Week 6–8 — "ChatGPT for Dummies" (~30–35 pages/day with hands-on tool practice alongside reading); Weeks 9–10 — review, workflow
- Human-AI collaboration as a force multiplier — Duffey's core argument that AI amplifies human creativity and judgment rather than replacing it, and how to position yourself as the 'superhuman' in the loop
- The IDEA framework from Superhuman Innovation — Inspire, Design, Evaluate, Accelerate — as a structured lens for integrating AI into creative and professional workflows
- Prediction machines and the economics of AI — Agrawal's central thesis from Power and Prediction that AI is fundamentally a cheap prediction technology, and what that means for how decisions get made at work and in life
- Power shifts from decision-makers to data — Power and Prediction's analysis of how AI redistributes organizational power, and why understanding who controls the prediction layer matters for navigating AI-driven workplaces
- Rule-based vs. judgment-based decisions — Agrawal's distinction between decisions that can be automated with AI predictions and those that still require human values and judgment
- Practical prompt construction for real tasks — Baker's hands-on guidance in ChatGPT for Dummies on writing effective prompts for writing assistance, research synthesis, brainstorming, and productivity tasks
- Iterative prompting and output refinement — the workflow of treating AI responses as drafts, using follow-up prompts to improve, constrain, or redirect outputs toward a specific goal
- Building a personal AI workflow — synthesizing all three books into a repeatable, personalized system for using AI tools daily across writing, research, and decision-support tasks
- According to Duffey's Superhuman Innovation, what is the human role in an AI-augmented creative or professional process, and how does the IDEA framework operationalize that role?
- Agrawal argues in Power and Prediction that AI is a 'prediction machine' — what does this mean in practical terms, and how does cheap prediction change the value of human judgment in your own job or daily decisions?
- Using the framework from Power and Prediction, can you identify one decision in your work or life that could be partially automated by AI prediction, and one that must remain human-led — and explain why?
- Based on Baker's ChatGPT for Dummies, what are the key structural elements of an effective prompt for a complex real-world task (e.g., drafting a report, summarizing research, or planning a project)?
- How do the perspectives of all three books align or conflict on the question of AI and human agency — does Duffey's optimism, Agrawal's economic realism, and Baker's pragmatism point to the same kind of human-AI relationship?
- What does your personal AI workflow look like after completing this stage — which tools, prompt patterns, and decision filters will you use regularly, and which book most influenced that design?
- IDEA Framework Audit (Superhuman Innovation): Choose one real project at work or in your personal life. Map it through Duffey's IDEA stages — write 2–3 sentences for each stage describing how AI could assist, and identify where your human judgment is irreplaceable.
- Prediction Mapping Exercise (Power and Prediction): List 10 decisions you make in a typical week. Classify each as 'AI-predictable,' 'human judgment required,' or 'hybrid.' Use Agrawal's rule-based vs. judgment-based framework to justify each classification, then reflect on whether your current tool use matches this map.
- Prompt Lab — Real Task Sprint (ChatGPT for Dummies): Using Baker's prompt-construction guidance, write and test prompts for five real tasks: (1) drafting a professional email, (2) summarizing a long article, (3) brainstorming solutions to a current problem, (4) creating a weekly plan, and (5) explaining a complex topic to a non-expert. Document what worked and what needed refinement.
- Iterative Prompting Diary: Pick one substantial writing or research task. Complete it using at least 4 rounds of iterative prompting — log each prompt, the output, what you changed, and why. Reflect on how the final output compares to what a single prompt produced.
- Power Shift Reflection (Power and Prediction): Write a one-page memo as if advising your organization (or yourself as a freelancer/individual) on how AI-driven prediction will shift decision-making power over the next 3 years. Draw directly on Agrawal's frameworks and propose one concrete adaptation.
- Personal AI Workflow Document: Synthesize all three books into a one-to-two page 'My AI Workflow' reference sheet — include your go-to prompt templates (Baker), your decision filter for when to use vs. not use AI (Agrawal), and your human-in-the-loop checkpoints (Duffey). Treat this as a living document you will refine in the next stage.
Next up: Completing this stage equips the reader with a working personal AI toolkit and a critical economic lens on AI decision-making, creating the practical foundation needed to tackle more advanced topics such as prompt engineering at scale, AI ethics, and designing AI-augmented systems for teams or organizations.
Bridges the gap between prompting skills and business/creative application, showing how to embed AI into actual workflows rather than using it ad hoc.
Reframes how AI changes decision-making at work and in life, helping you identify where AI assistance adds the most leverage in your own context.
A practical, task-by-task reference for using ChatGPT across writing, research, coding, and communication — ideal for cementing everyday habits at this stage.
Advanced Prompting & AI Thinking
Going deepMaster advanced prompt patterns, understand the reasoning and limitations of LLMs well enough to work around them, and think critically about AI outputs like an expert user.
▸ Study plan for this stage
Pace: 6–8 weeks total: ~3 weeks on "Impromptu" (~25–30 pages/day, including annotation time) and ~3–4 weeks on "A Hacker's Mind" (~20–25 pages/day, given its denser conceptual load). Reserve the final 3–4 days for cross-book synthesis and exercise completion.
- Advanced prompt patterns: chaining, role-playing, constraint-setting, and iterative refinement as demonstrated through Hoffman's co-creative dialogues with GPT-4
- Collaborative intelligence: treating LLMs as a 'thought partner' rather than a search engine — shaping outputs through framing, context, and follow-up
- LLM reasoning boundaries: understanding where models hallucinate, over-generalize, or produce confident-sounding errors, and how to design prompts that surface these failure modes
- Hacking as a cognitive lens (Schneier): identifying and exploiting gaps between the intent of a system and its actual rules — applied to AI systems and their prompt interfaces
- System vs. rule manipulation: how advanced users 'hack' LLM behavior by finding edge cases in training assumptions, RLHF guardrails, and instruction-following patterns
- Critical evaluation of AI outputs: distinguishing fluency from accuracy, recognizing sycophancy, and stress-testing model claims with adversarial follow-up prompts
- Societal and power implications: how AI capabilities concentrate leverage among those who can prompt effectively, echoing Schneier's analysis of who benefits from system exploits
- Ethical red-teaming: using a hacker's mindset responsibly — probing AI limitations to improve outputs without causing harm or circumventing safety measures
- After working through Hoffman's dialogues in 'Impromptu', how would you construct a multi-turn prompt sequence to explore a complex, ambiguous professional problem — and what iterative adjustments would you make when the first response misses the mark?
- Schneier argues that 'hacking' is fundamentally about finding gaps between intent and implementation. How does this framework apply to prompt engineering, and what specific LLM behaviors does it help explain?
- What are the most common failure modes of LLMs (hallucination, sycophancy, context drift) and how can prompt design — informed by both books — be used to detect and mitigate each one?
- Hoffman frames AI as a collaborative amplifier of human expertise. How should an expert user calibrate how much creative or analytical authority to delegate to an LLM versus retaining for themselves?
- Using Schneier's concept of 'hacking social systems', how might you critically evaluate whether an AI-generated argument, plan, or piece of content is subtly optimizing for something other than your stated goal?
- How do the power dynamics Schneier identifies in system exploitation map onto the gap between casual AI users and expert prompt engineers — and what are the ethical responsibilities that come with advanced prompting skill?
- Prompt chain lab (Impromptu-based): Pick a real strategic or creative challenge you face. Run a minimum 6-turn dialogue with an LLM, deliberately applying one new constraint or role-shift per turn (e.g., 'now argue the opposite', 'respond as a skeptical domain expert'). Document what changed and why.
- Failure-mode hunting: Design 5 prompts specifically intended to surface hallucination, sycophancy, or overconfidence in an LLM. For each, write a 'detection prompt' — a follow-up that stress-tests the original answer. Record results and refine your detection techniques.
- Hacker's audit (Schneier-based): Choose one AI tool you use regularly. Write a one-page 'system audit' identifying: (a) the tool's intended rules, (b) gaps between intent and implementation you've observed, and (c) two 'exploits' — prompt strategies that get better results by working around default behavior rather than with it.
- Sycophancy stress-test: Submit a prompt containing a subtly flawed argument or factual error. Observe whether the model agrees, flatters, or pushes back. Then redesign the prompt with explicit instructions for adversarial review. Compare outputs and write a brief analysis.
- Cross-book synthesis essay (500–700 words): Write a personal framework titled 'How I Think About AI Outputs' that integrates Hoffman's collaborative intelligence model with Schneier's exploit-and-audit mindset. Include at least three concrete prompting heuristics you will use going forward.
- Red-team a real output: Take an AI-generated piece of content (an email, a plan, a summary) — either one you produced or a public example. Apply Schneier's hacker lens to ask: 'What is this output actually optimizing for?' Annotate it line by line, then write an improved prompt that closes the gaps you found.
Next up: By internalizing both the collaborative craft of advanced prompting (Hoffman) and the critical, systems-level skepticism of a hacker (Schneier), the reader is now equipped to move beyond using AI effectively in isolation — and into evaluating, designing, and governing AI workflows at a broader, more consequential scale.
Written in deep collaboration with GPT-4, this book models sophisticated, high-level prompting in action and explores the frontier of what AI can do as a thinking partner.
Teaches you to think adversarially and creatively about systems — including AI — so you can probe, test, and push your prompts beyond surface-level use and spot when AI misleads you.