Prompt Engineering Full Course

Summary of “Prompt Engineering Full Course” Subtitles (Tech Concepts & Techniques)

Bad vs. Good Prompting

• Bad prompt: “write something about our product”
  • Tends to produce generic marketing fluff
  • Often gets tone and length wrong
  • May miss a clear CTA
• Good prompt: specifies details such as:
  • Role (e.g., senior B2B copywriter)
  • Task (e.g., a 2-sentence LinkedIn ad)
  • Product context (e.g., project management SaaS)
  • Audience (e.g., ops managers at midsize companies)
  • Tone (e.g., confident, not salesy)
  • Output requirement (e.g., ends with a clear CTA)

Key point: The improvement comes from instructions, not from changing the LLM.

---

What Prompt Engineering Is (and Why It Matters)

• Treated like programming in natural language: you “code” by writing text/voice instructions rather than Python/Java/etc.
• LLMs don’t have a built-in “task list,” so prompts must explicitly define:
  • task
  • role
  • format
  • constraints
• Models are moving toward an agent era:
  • LLMs can call tools (search, create Docs, write slides, take actions)
  • Prompting therefore affects not just text output, but actual actions

---

Speeding Up Prompts via Dictation

• Recommends speaking prompts instead of typing to include more detail faster.
• Uses Whisper Flow:
  • Handles punctuation and capitalization
  • Removes filler words
  • Supports bullet formatting
  • Includes features like dictionaries/snippets, app-style switching, keyboard shortcuts
  • Works on Android and phone keyboards
• Motivation: faster iteration helps produce better prompt quality and boosts productivity

---

How LLMs “Think” (Core Mental Model)

• Fundamentally a text prediction model (input tokens → output tokens)
• By default, it has no real memory
  • “Memory” is usually context injection by the interface/tools
• Many interfaces add hidden context, such as:
  • previous messages
  • tool access
  • system prompts Meaning the model may “see” more than what you directly typed.

---

Steering vs. Commanding

• Commanding: “summarize this”
  • Model chooses length/style/focus
• Steering: provides explicit direction, e.g.:
  • “You are an executive assistant… summarize in four bullet points… focus on decisions and action items… no filler.”

Steering improves accuracy by constraining:

• format
• length
• focus
• exclusions (“what not to do”)

---

Core Prompting Techniques Taught (with Examples)

1. Specificity / Set the Scene

• Always better to include: role + audience + tone + output format
• Example (rewriting a support reply):
  • “reply to this complaint” → generic results
  • Adding role (“customer support lead”), the situation, constraints (under 150 words), and sign-off → more useful output

---

2. Prompt Formatting & Clear Structure

• Use bullet points
• Separate sections clearly
• Use delimiters like: “here is the complaint: …”

---

3. Few-shot Prompting

• Provide example input-output pairs so the model learns the pattern.
• Useful for:
  • classification tasks
  • consistent structured outputs (e.g., ticket titles)
• Example: producing a specific one-line ticket title format
  • Improved once examples + format constraints were added

---

4. Chain-of-Thought Prompting

• Encourage step-by-step reasoning to reduce:
  • logic errors
  • math/planning errors
• Note: modern “reasoning/planning” models may do this automatically, but some APIs may require explicit instructions.

---

5. Structured Output

• Request machine-parseable output like JSON
• Provide an example schema/shape
• Example: comparing Trello/monday.com/clickup and outputting valid JSON only → makes it usable directly for APIs or databases.

---

6. Constraints & Negative Instructions

• Sometimes the best prompts emphasize what not to do.
• Constraint examples include:
  • exact length (“under 300 words”)
  • tone (“no slang or humor”)
  • do/don’t rules (e.g., “Do not apologize”)
  • formatting rules (“Do not use bullet points”)
• Example: an out-of-office email improved by specifying exact dates/contact info and forbidding fluff.

---

7. Iterative Refinement

• Prompting works like a loop:
  • the first output is rarely perfect
  • then refine by requesting changes (shorter/more formal/add examples/focus on X)
• Prefer incremental improvement over restarting from scratch.

---

8. Interview-Style Prompting (Underrated)

• Instead of guessing context, have the model interview you:
  • ask one question at a time
  • gather details
  • generate the final output after enough information is provided
• Example: creating a ~250-word LinkedIn post about switching to a 4-day work week
  • The model asked clarifying questions about audience, company context, goals, metrics, and do/don’t rules
  • Result: more accurate and less generic

---

Advanced Strategies

System vs. User Prompts

• System prompts: persistent behavior/rules/style (often hidden from end users)
• User prompts: what you type for a specific task
• Mentions ChatGPT “custom instructions” as system-level guidance.

---

Prompt Chaining (Multi-step Workflows)

• Break complex tasks into sequential steps:
  • output of step N becomes input for step N+1
• Improves reliability compared to a single large prompt.

---

Self-evaluation

• Ask the model to:
  • critique/score its output
  • propose improvements
• Tip: reduce bias by refreshing context (e.g., “I wrote this” vs. “an AI wrote this”) and not always evaluating in the same chain.

---

Temperature & API Parameters

• Temperature controls determinism:
  • Lower → more repeatable/consistent
  • Higher → more varied/creative
• Suggested usage:
  • low temp for structured answers/code/facts
  • higher temp for brainstorming or multiple phrasing ideas

---

Common Mistakes + Fixes (as Taught)

• Too vague → add role/audience/tone/format/length.
• Too many tasks in one prompt → split into steps/prompts (use chaining).
• Not enough context/examples → add 1–3 shot examples or use interview-style prompting.
• Model ignores required format → explicitly demand the format and “output only” (e.g., JSON only).
• Assuming memory → repeat key facts when starting new sessions or if context window limits are reached.

---

Key Sources / Speakers (from Subtitles)

• Main speaker/creator: course instructor/author presenting prompts in ChatGPT and Cursor
• Mentioned third-party tool: Whisper Flow (dictation tool)