Develop AI competencies through structured practice, real evaluation frameworks, and deliberate technique. Move from casual use to confident, critical mastery.
Everyone can type a question into an AI. Mastery is the repeatable ability to shape, evaluate, and reason with its output. These three skills compound.
The craft of designing instructions that reliably produce the output you need. Structure, context, examples, constraints, and iteration, not magic words.
A systematic way to judge AI responses for accuracy, bias, completeness, and appropriateness to your task. Trust is earned per output, never granted by default.
Using AI to sharpen your reasoning without outsourcing it. Knowing when to accept, when to push back, and when to set the tool aside entirely.
No tracking, no accounts. Takes two minutes and gives you a realistic read on your current level and where to focus next.
Filter by your role above to see examples tuned to your work. Filter by level to jump to where you are.
You've used AI before, but results feel inconsistent. The goal at this level is to reliably get a response that's in the right ballpark without starting over five times.
You get decent responses, but complex tasks still take many rounds. The goal here is technique: you learn specific moves that consistently improve results.
You use AI fluently, now you need discipline. The goal at this level is judgment: knowing when to lean in, when to pull back, and how to keep your own voice.
Each technique is a specific, named move. See it fail, see it work, and understand why.
Most "bad AI output" is a response to an ambiguous prompt. Models pattern-match to the most likely interpretation of vague input. Specifying your task, audience, and desired format resolves that ambiguity up front.
The second prompt fixes four things: the audience (so the vocabulary is right), the length (so the model doesn't over-deliver), the structure (so key points aren't missed), and the style (so it's concrete instead of abstract).
Assigning a role shapes vocabulary, default assumptions, and depth of explanation. It's not about trickery, it's about specifying the voice and expertise level you want the response delivered in.
The role (academic writing tutor) narrows the feedback to what matters for the writer. The three-point structure prevents a generic "here are ten things" list and forces prioritization.
Describing a style or format in words is imprecise. Showing examples is dramatically more precise. Few-shot prompting is the single biggest quality upgrade for any task with a consistent structure (summaries, rewrites, classifications, formatting).
The examples teach the model the exact shape, tone, and granularity you want. Patterns communicate what instructions can't.
On reasoning tasks (math, logic, multi-step analysis), asking the model to explain its thinking step-by-step before giving a final answer measurably improves accuracy. You also get a reasoning trace you can check for errors.
Forcing the model to externalize intermediate steps reduces the chance of skipping a computation, and lets you catch an error at step 2 before it contaminates step 4.
A single mega-prompt asking for analysis + writing + formatting + citations will produce mediocre output in all four dimensions. Running each step as its own prompt, with the previous output as input to the next, produces noticeably better results with clearer points to verify.
You verify at each step. Errors can't compound silently. And each sub-task is simple enough that the model performs well on it.
Negative constraints ("don't use jargon", "no introduction", "no more than 3 bullets") are often more useful than positive ones. Format constraints (exact length, specific structure, required sections) keep the output usable instead of a wall of prose you then have to re-shape.
Constraints prevent the model's default tendencies (over-apologizing, over-explaining, generic sympathy) and protect your voice. The "do not mention reasons I haven't specified" line alone rescues most of these emails.
Novice users accept the first output or start over from scratch. Skilled users refine. Ask what's weak. Ask for the strongest version of a specific point. Ask for three alternatives to a passage you don't like. The best prompts are almost always the third one.
"Make it better" has no target. Specific, localized revision requests with concrete replacement instructions give the model something to aim at. And asking for alternatives lets you choose instead of edit.
Use this the first few times you review AI work. After a month it becomes automatic, a habit instead of a checklist.
Self-guided, no setup. Three challenges per level, each mapped to a core competency. Use the tabs below to switch levels.
Take a prompt you wrote in the past week. Rewrite it three times: once adding context, once adding format, once adding constraints.
Which version produced the biggest quality jump? Why? Which rewrite would you always apply from now on?
Ask an AI for five specific facts in a niche area of your major. Verify each one independently. Track how it fails.
Which kinds of errors were easiest to spot? Which were most convincing? What heuristic would you use next time to catch similar errors faster?
Take a position you actually hold. Ask the AI for the three strongest counterarguments. Actually engage with them.
Did the AI generate challenges you hadn't considered, or mostly predictable ones? Where did it help you think, and where did it feel like productivity theater?
Take a complex task you'd normally write one prompt for. Break it into 5 sequential steps. Verify at each handoff.
Where in the chain did you catch something you would have missed? Was the extra time worth it? For which task types will you use this pattern by default?
Run the same non-trivial task on two different AI tools. Evaluate against a rubric YOU write before you see either output.
Did your rubric capture what actually mattered, or did you find yourself wanting criteria you hadn't written down? What does that tell you about your implicit standards?
For seven days, refuse to use AI for one specific type of task you normally rely on it for. Journal the change.
What did you lose? What did you find? Where had AI been genuinely helping, and where had it been doing the thinking that was supposed to be yours?
Have AI summarize a topic, paper, or book you know deeply. Compare line by line to what you actually know and catch every distortion.
Where did AI confidently say something wrong? Where did it capture the surface but miss what actually mattered? What does that pattern tell you about trusting AI summaries on topics you don't already know well?
Write your own analysis before showing AI. Then get AI's pushback. Compare to the reverse order on a separate question.
Which order produced clearer thinking? Which produced better final output? Where did AI's pushback genuinely sharpen your view, and where did it just add noise you ended up dismissing? When should you use each pattern?
Create a prompt template for a task you do repeatedly. Stress-test it against 5 genuinely different inputs. Iterate until it holds.
What kinds of inputs caused the most revision? What was the template unable to handle even after iteration? At what point did you stop improving it, and why? Is this template now worth keeping permanently, or did testing reveal that the task isn't actually suited for templating?
Hand-picked, no course aggregators or marketing pages. Every link opens to a real resource, verified for this page.
Nine-chapter step-by-step course with hands-on exercises and an answer key. Free, self-paced, free-to-clone.
Andrew Ng and Isa Fulford (OpenAI) teach principles of effective prompting with working notebook examples. Free.
Jules White's patterns-based course on prompt patterns, persona prompting, and few-shot techniques. Auditable for free.
Six hours covering what LLMs are, how to use them well, and how to evaluate outputs. Ideal before deeper technique courses.
Interactive environment to test prompts with Claude, adjust system prompts and parameters, and see how settings change output.
Experiment with GPT models, system messages, temperature, and response format. Useful for comparing behavior to Claude.
Test prompts against Gemini models with long-context support. Handy for trying the same prompt across three major providers.
Run the same prompt on two models simultaneously in blind comparison. The fastest way to build intuition for which tool suits which task.
Comprehensive open-source reference covering zero-shot, few-shot, chain-of-thought, self-consistency, tree-of-thought, and more.
Anthropic's official guide to prompt design, covering clarity, examples, chain-of-thought, XML structure, and iterative refinement.
Academic paper cataloging reusable prompt patterns (persona, template, flipped interaction, game-play). Foundation for pattern-based thinking.
Practical, honest, and frequently updated writing on prompt injection, prompt design, and evaluation by a leading independent practitioner.
Peer-reviewed research on evaluating language models, including the Foundation Model Transparency Index and HELM benchmark analyses.
Deep technical overview from an OpenAI researcher. Good after the basics, covers self-consistency, automatic prompt generation, and evaluation.
The techniques on this page transfer across every AI Hub page. Pick the domain that matters most for your next step.
If any of this felt too far ahead, step back and build from the ground up. Concepts, starter tools, responsible use.
Go to New to AI →Apply prompt engineering and evaluation to course design, assessment, and student support. Pedagogy-specific patterns.
Go to Teaching with AI →Use these techniques for literature review, analysis, and writing. Research-specific workflows and evaluation.
Go to Research with AI →