Levels 1 & 2

AI Techniques: 76 ways to control AI

From basic prompting to multi-agent architectures. The techniques behind semantic functions and AI agents. Each one is a tool you can use today.

Technique (34)

Ask a Better QuestionLet AI improve your question before answering it. A better question gets a better answer.Ask for OptionsDon't settle for one answer. Ask AI for multiple approaches and pick the best one.Ask It to SearchAI can search the internet now. You just have to ask.Be SpecificVague questions get vague answers. The more details you give, the better the result.Break Down the QuestionComplex questions need smaller questions first. AI reasons better when problems are split Challenge MeAI wants to agree with you. Ask it to push back instead.Complexity-Based PromptingWhen giving AI examples, pick the richest, most detailed ones. Complex examples teach thor Contrastive Chain-of-ThoughtShow AI what good reasoning looks like — and what bad reasoning looks like. It learn Extract What MattersGive AI messy content — a long email, a photo, a wall of text — and ask it to Generate Examples FirstAsk AI to recall similar problems before solving yours. Its own examples guide better reas Give It a RoleTell AI who to be. A specific role shapes better, more useful answers.Give It the SourceDon't ask AI to work from memory. Paste the actual document, and AI works from facts inste Interview MeLet AI ask the questions first. You'll clarify your own thinking — and AI will know IterateDon't expect AI to get it right on the first try. Evaluate, give feedback, repeat.Lead With Your IdeaDon't ask AI to create from nothing. Give it something of yours to work with.Let It Say I Don't KnowAI will guess before it admits uncertainty. Unless you tell it not to.Make It FamiliarAsk AI to explain new concepts using something you already understand — your hobbies Make It MatterAdd emotional stakes to your prompt. AI tries harder when it feels important.One Thing at a TimeBig prompts with multiple tasks get mediocre results. Break it up.Plan FirstFor complex tasks, ask AI to outline the steps before doing anything.Recall FirstAsk AI what it knows before asking what you need. The right context leads to better answer Set ConstraintsTell AI where to stop. Boundaries keep answers focused and prevent unwanted extras.Set the FormatTell AI how to structure the output. The same information becomes immediately usable.Show by ExampleDon't just describe what you want. Show it. Examples teach AI faster than instructions.Show ItSome things are hard to describe in words. Upload an image instead — AI can see it.Show the SourcesAsk AI to list the key facts it used. You'll know exactly what to verify.Structure the OutputAsk AI to return answers in a specific format — like JSON — so other tools and System 2 AttentionStrip out the noise, opinions, and bias from your context before AI answers. Clean input, Tell It Who It's ForThe same topic explained to a beginner looks nothing like the expert version. Tell AI who'Think Step by StepAsk AI to show its reasoning. Better thinking leads to better answers.Thread of ThoughtWalk through long or messy context piece by piece before answering. Nothing gets missed.What Am I Missing?Ask AI to find your blind spots. The things you can't see because you're too close.You DriveAI is not always right. You decide what's good.Zoom Out FirstAsk for the general principle before the specific answer. AI reasons better when it sees t

Composition (34)

Active PromptingFind the questions AI struggles with most, then teach it exactly those. Spend your effort APE – Automatic Prompt EngineerAutomatic Prompt Engineer. Instead of hand-crafting prompts through trial and error, let A Chain ItFeed the output of one prompt into the next. Build complex results step by step.Chain of DensityStart with a loose summary. Then pack in more information — without adding more word Chain-of-ActionBefore answering each sub-question, check: "Do I actually know this?" If not, pause reason Check Your WorkAI often glosses over details. Ask it to verify its claims — it will frequently catc Critique and ReviseAsk AI to evaluate its own output, then improve it. A simple loop that polishes writing.Cumulative ReasoningPropose a step. Verify it. Only keep what checks out. Build your answer on confirmed facts Decomposed PromptingBreak the task into pieces. Send each piece to the specialist best equipped to handle it.Define Your ToolsTell AI what functions it can call. Good tool definitions are the key to reliable function Directional Stimulus PromptingA small, trainable model generates targeted hints — keywords, key points — tha DSPyStop writing prompts. Declare what you need, build it like software, and let a compiler fi Give It Your ToolkitTell AI what tools you have — functions, APIs, services — and let it assemble Index FirstShow the map before the territory. Let AI choose what context it actually needs.Least-to-Most PromptingBreak the problem into sub-problems. Solve the easiest first. Feed each answer into the ne Let Code Do ItInstead of asking AI to solve a problem directly, ask it to write code that solves it. The LLMCompilerMap your task into a graph of dependencies. Fire off independent steps in parallel. Cut la Loop Until DoneKeep iterating until a condition is met. Let AI refine its own work automatically.Maieutic PromptingQuestion AI's beliefs from both sides. When explanations contradict each other, the truth Meta-PromptingOne AI acts as a conductor, creating specialized expert personas on the fly and delegating Multimodal Chain-of-ThoughtFirst describe what you see. Then reason about what it means. Two stages that turn visual Plan-and-ExecuteMake the full plan first. Then work through it step by step. Separate thinking from doing.Program of ThoughtsLet AI reason about the problem and write the logic. Let a code interpreter handle the mat RAG PatternsSearch your knowledge base first. Then answer grounded in real documents. The AI stops gue ReActThe foundational agent loop. Think about what you need, use a tool to get it, look at what Recursive Chain-of-FeedbackDon't just say "try again." Find the exact step that went wrong, isolate it as a simpler p ReflexionWhen AI fails, don't just retry. Have it reflect on what went wrong, remember the lesson, ReWOOPlan all the tool calls upfront using placeholders. Run the tools without AI. Then synthes Route ItClassify first, then send to specialized handlers. Different inputs get different treatmen Self-AskAI generates its own follow-up questions, answers each one, then combines them into a fina Self-ConsistencyAsk AI the same question multiple times. The answer that shows up most often is usually th Skeleton of ThoughtOutline first, then expand every point at the same time. Faster results, better structure.Stack ThemCombine multiple techniques into a single powerful prompt. Layer role, audience, examples,Toolformer & TALMTeach AI to recognize when it needs a tool and call it naturally — mid-sentence, wit

System (8)

Adaptive Pattern RouterThe dispatcher that sits above every technique. Classify the task, pick the best approach,AutoGPT / BabyAGIGive the AI a goal and let it loose. It creates its own task list, prioritizes, executes w Cognitive LoopSeven stages. One complete thinking system. Perceive, Retrieve, Reason, Plan, Act, Verify,Generative AgentsAI characters that remember, reflect, and plan. A memory stream architecture that produces JARVIS / HuggingGPTOne AI brain orchestrating many specialist models. It plans the workflow, picks the best m LATSLanguage Agent Tree Search. Explore multiple solution paths like a chess engine — us Multi-Agent CompositionsOne AI with many roles is limited. Multiple specialized agents — researcher, analyst VoyagerA lifelong learning agent. It teaches itself new skills, verifies they work, and stores th