Nearby lessons

46 of 49

Prompt Engineering - Quick Cheatsheet

Chapter 1 — Introduction to Prompt Engineering

  • A prompt is an instruction given to AI; prompting is a communication skill.
  • AI is a pattern follower, not a mind reader — clear instruction gives good output.
  • The auto-driver example: a clear destination is half the journey.
  • How AI works: type → read → understand → generate.
  • Good prompt = specific + complete + tells audience and purpose.
  • Bad prompt = vague + missing details → AI guesses.

Chapter 2 — Prompt Clarity Rules & the Prompt Formula

  • Prompt clarity = specific + complete — missing either makes AI guess.
  • Rule 1 — Be Specific: vague words like "explain this" waste a turn.
  • Rule 2 — Be Complete: "Write resume" gives a generic resume; add skills and experience.
  • The prompt formula has 5 parts: Role, Task, Context, Constraints, Output format.
  • Role gives direction, context gives clarity, constraints give control.
  • The tailor example: a structured instruction gives a perfect result.
  • Apply the formula to every prompt and AI output becomes professional.

Chapter 3 — Tone, Length & Step-by-Step Control

  • Tone controls how AI speaks; same topic, different style.
  • Four tones: Friendly (learning), Professional (office), Funny (content), Strict (interview).
  • Length control decides the size: short for revision, long for deep learning.
  • Step-by-step prompting gives clear, ordered explanations.
  • Combine tone + length + steps in one prompt for the best result.
  • Practice daily — real prompting skill comes from doing, not reading.

Chapter 4 — Format Control & Structured Output

  • Format control decides how the answer looks, not what it says.
  • Five formats: bullets, table, steps, checklist, story — each with a best use.
  • Structured output = fixed headings + strict tables = predictable, reusable output.
  • Without "strictly", AI may quietly change the headings or add its own sections.
  • "Only Output This" removes extra text and makes AI tool-ready (JSON, APIs).
  • Only-output is essential for JSON, apps, and APIs — extra text breaks the data.
  • The notes-generator prompt: role + task + fixed structure + rules + only-output.
  • Same input → same format → easy reuse and automation.

Chapter 5 — Small Words, Constraints & Ambiguity

  • AI reads prompts step by step; instruction order matters — put role first.
  • Last instructions dominate, like the WhatsApp meeting example.
  • Control words — only, exactly, strictly, do not, must — give word-level precision.
  • One small word can change the length, tone, or format of the entire output.
  • Ambiguity makes AI guess; a prompt should read like a requirement document.
  • Constraints are rules that limit AI behavior — length, tone, format, and plain rules.
  • An example is a sample output that guides AI's style — a visible reference.
  • Constraints + examples = control + stability.

Chapter 6 — Zero-Shot, One-Shot & Few-Shot Prompting

  • Zero-shot: ask directly, no example — fast but not controlled.
  • One-shot: one example locks the style, format, and tone.
  • Few-shot: 2–5 examples teach AI a pattern and improve accuracy.
  • Few-shot works best for classification, MCQs, and email replies.
  • Few-shot + structure = production-ready, reusable AI output.
  • Teacher analogy: 1 example = little, 3 examples = clear, 5 examples = mastery.

Chapter 7 — Persona, Role & Tone Control

  • Role = job, Persona = behaviour, Tone = voice — three different things.
  • Persona control = behaviour + depth + attitude.
  • Strict interviewer, mentor, analyst, HR — each persona changes the output.
  • Persona + constraints + format = the stability layer for reusable AI tools.
  • Domain persona makes AI think from a field's perspective.
  • Safety persona refuses harmful requests and suggests safe alternatives.

Chapter 8 — Chain-of-Thought & Reasoning Foundations

  • A reasoning prompt makes AI think step by step before answering.
  • "Think step by step" is a small line with a big effect.
  • Manual CoT: you write the thinking steps; full control.
  • Auto-CoT: AI generates the reasoning examples; faster.
  • Showing its working makes the answer believable — and trustworthy.
  • Not every question needs step-by-step reasoning — simple questions get quick answers.
  • Long prompts for simple questions waste time and tokens.
  • Efficiency = depth when needed, speed when not.

Chapter 9 — Advanced Reasoning: Self-Consistency, ToT & Step-Back

  • Reflection = AI reviews and improves its own answer.
  • Self-consistency = solve multiple times and choose the common answer.
  • Tree of Thoughts = generate options, compare, decide like a leader.
  • Chain of Hindsight = learn from mistakes and improve.
  • Step-Back = understand the concept before solving.
  • Check twice, answer once — combine reflection + consistency for trust.

Chapter 10 — Graph of Thoughts & Reliability Engineering

  • GoT thinks in a connected network — one change affects many things.
  • CoT is linear, ToT is branching, GoT is a full system.
  • Reliability engineering = safe + accurate + verified AI.
  • Guardrails control allowed/not-allowed topics.
  • Hallucination control stops AI from creating fake facts.
  • Complex simulation combines CoT + ToT + Self-Consistency + Reflection.

Chapter 11 — ReAct, Tool Calling & Stop Conditions

  • ReAct = Reason + Act, repeated until the goal is done.
  • Every step alternates: Reason (think) then Act (do) — the loop is the foundation of AI agents.
  • AI becomes an assistant that completes tasks, not just an advisor.
  • Tool calling = calculator, search, database, validator — verify, don't guess.
  • Use tools when accuracy matters; skip them for simple answers.
  • A stop condition tells AI when to stop — without one it can loop forever.
  • Stop conditions + retry limits + output limits = safe agents.
  • Smart AI knows when to stop.

Chapter 12 — Planning Prompts & Mini AI Agents

  • A planning prompt turns a big goal into phases, tasks, priority, timeline, milestones.
  • Order matters — buy land → design → build → paint; wrong order = failure.
  • A milestone is a big checkpoint that measures progress.
  • An AI agent thinks + acts + verifies + improves as a full workflow.
  • Mini agent = CoT + ToT + CoH + Reflection + ReAct + guardrails + stop.
  • One single prompt combines every technique you learned.
  • The capstone combines all your skills into one real-world system.
  • Capstone = from learner to professional.

Chapter 13 — Debugging Prompts: Failures, Refactoring & the Meta-Prompt

  • AI is not failing — the prompt is failing.
  • Prompt failure = wrong or unclear output because the instruction was poor.
  • Refactoring = same meaning, clear structure, no noise — break the prompt into role, context, constraints, output format.
  • Debugging process: Diagnose → Explain → Rewrite → Test.
  • A good prompt is written, a great prompt is tested — checklist: role, task, format, ambiguity, missing data, testability.
  • Improvement loop: Ask → Check → Refine → Retest.
  • A meta-prompt works on another prompt — analyzes, improves, and rewrites it.
  • The next level is not using AI, it is building AI.

Chapter 14 — Decomposition & Linear Pipelines

  • One big prompt fails for complex tasks — chain small prompts instead.
  • Chaining: output of one step becomes input of the next.
  • Decomposition uses Input → Process → Output for every step.
  • A linear pipeline runs steps in a fixed order, each depending on the last.
  • Format locking keeps the same structure across all steps.
  • Notes → summary → quiz is the classic beginner pipeline.

Chapter 15 — Branching Pipelines, Controlled Chaining & Variable Injection

  • A branching pipeline changes path based on an IF condition — IF → Route A, ELSE → Route B.
  • Different users → different outputs → personalization.
  • Classify the user first, then send them down the correct path.
  • Controlled chaining keeps the same topic, format, and rules across steps.
  • "Do not add new information" prevents topic drift; each step uses only the previous output.
  • Variable injection uses placeholders like {{TOPIC}} to make prompts reusable.
  • Identify the dynamic parts, replace them with variables, reuse the template.
  • Static prompts are one-time; dynamic templates scale.

Chapter 16 — Reusable Templates, Automation & the Capstone

  • Reusable templates = modules + variables + standard structure.
  • Automation shifts you from prompts to systems.
  • JSON passes structured data between steps automatically.
  • The capstone combines chaining, pipelines, variables, and structure.
  • Optimization improves speed, accuracy, cost, and stability — remove extra steps, reduce tokens, finalize the template.
  • Build once, use many times — creators build systems.

Chapter 17 — Prompt Formats: System/User, Alpaca, ChatML & INST

  • Prompt serialization = structuring prompts with roles.
  • System message controls behaviour; user message gives the task.
  • System message has higher priority.
  • Alpaca = Instruction + Input + Output (model training).
  • ChatML = System / User / Assistant with context memory.
  • INST wraps instructions in [INST] tags for LLaMA models.

Chapter 18 — Minimalism, Tokens, Cost & Safety

  • Minimalism = short + clear + complete.
  • Noise removal = cut words that add no value.
  • A token is a piece of text; more tokens = more cost.
  • Short prompts = low tokens = low cost.
  • Consistency = same style and format every time.
  • Safety: ask assumptions, sources, verification; allow "Not sure".

Chapter 19 — Multimodal, Image & Video Prompting

  • Multimodal AI handles text, image, video, and audio together.
  • Image formula: Subject + Environment + Lighting + Camera + Style.
  • Lighting controls atmosphere, camera controls perspective, mood controls emotion.
  • Video formula adds Action + Camera Movement to the image formula.
  • Detailed prompts give professional, cinematic output.
  • Transformation videos are powerful for viral content.

Chapter 20 — Negative Prompting & PAL Prompting

  • Negative prompting tells AI what NOT to generate.
  • Prompt = what you want; negative prompt = what you don't want.
  • Reuse the negative word list: blurry, extra fingers, watermark, text, logo...
  • PAL = ask AI to write a program and solve, not to guess an answer.
  • PAL formula: "Write a Python program to + [problem]".
  • PAL works for finance, real estate, data, and automation — not for stories.
📝 Key Takeaways
  • One section per chapter, built from the chapter summaries
  • Each point is the exam-ready one-liner for that topic
  • Open the matching lesson for the full explanation and prompts