Prompt Engineering Guide

Sample prompt structures, types, and techniques to try

Concept Map
Prompting
Class Recording

Prompt Engineering Term

Prompt engineering was coined because prompts can vary widely and can be very simple or highly engineered. The quality of the input (prompt) greatly affects the quality of the output, comprehensiveness, and consistency. These three factors are the basis of why not all prompts are created equal and why any automation, both personally and for larger groups, requires prompt engineering.

Prompt Engineering Styles

Different prompting styles (Plain Text / Natural Language, Markdown / Structured Text, Templates, Code-Based Prompts)

Summary of major LLM models, uses, and differences

🤖

ChatGPT Models

ChatGPT has several models available. GPT is good for general use/questions, inference/qualitative semantic data, situational and personal preferences, context and tone modulation.

GPT5

A traffic cop that routes you to the optimal models for answers. This is both good and bad: good because you don't have to know the dozens of models OpenAI offers, but bad because you cannot choose a model that is best suited to your task and it may route you to the minimum viable model rather than the optimal one.

ChatGPT 4o legacy

Adopting GPT5 styling but with some cajoling can provide more in-depth, less fluffy answers. This will be addressed in the Tone part of the 101 course.

Access all GPT models

If you have a tool that uses an API connection, you can access all GPT models.

🧠

Claude Models

Claude will index on being correct rather than tailored to individual situations, making it ideal for analysis, quantitative data, scientific research, and high clarity.

Claude Opus 4.1

Most capable model for complex analysis, coding, and creative tasks. Excels at nuanced reasoning, lengthy document analysis, and maintaining context over long conversations. Best choice when accuracy and depth matter most.

Claude Sonnet 4

Balanced model offering strong performance at faster speeds. Ideal for most business tasks, content generation, and analysis where extreme depth isn't required. Cost-effective for production use.

Claude Haiku

(if available) Fastest, most cost-effective for simple tasks, quick responses, and high-volume processing. Perfect for straightforward queries and basic automation.

💪 Claude Strengths:
  • ✓Superior at maintaining consistency across long documents
  • ✓Better at following complex, multi-step instructions
  • ✓Excellent for technical documentation and code
  • ✓Strong ethical reasoning and safety considerations
  • ✓More likely to acknowledge uncertainty rather than hallucinate

Good Prompts vs. Bad Prompts

The difference between effective and ineffective prompting often comes down to specificity, context, and clear instructions.

❌

Bad Prompt Examples:

"Make it better" -- Vague, no context about what "better" means

"Write about AI" -- Too broad, no specific angle or purpose

"Fix this" -- No indication of what needs fixing or desired outcome

"Summarize" -- Missing context about length, focus, or audience

✅

Good Prompt Examples:

Product Description: "Rewrite this product description to emphasize sustainability benefits, targeting eco-conscious millennials, maintaining a friendly but professional tone, 150 words max"

Financial Analysis: "Analyze this financial report and identify the top 3 risks to profitability, providing specific data points and suggesting mitigation strategies for each"

Code Debugging: "Debug this Python function that should calculate compound interest but returns incorrect values. Focus on the mathematical logic and variable scoping"

Executive Summary: "Create a 5-bullet executive summary of this 50-page report, highlighting key metrics, decisions needed, and timeline constraints"

🔑

Key Elements of Good Prompts:

→
Context: What is the situation or background?
→
Specificity: What exactly do you want? Include constraints and requirements
→
Format: How should the output be structured?
→
Tone/Style: What voice or approach should be used?
→
Do's and Don'ts: What to include or avoid

Primary Uses / Types of Prompts

💬

General use queries

Add detail and structure output. Often tone modulation (to be covered in next session) is the guiding force here.

📊

Analysis / Research

To summarize, synthesize, and extract insight, which is helpful when doing research AND when reviewing content. High-value use cases are to refine a large dataset and to analyze/critique your own output.

🎨

Creative Output & Image Gen

Copywriting, storytelling, ad drafts, naming, branding. Often good to build a custom LLM with preset tone & voice -- this is good to organize thoughts. Image gen is similar but requires more one-offs with refinement of structured input.

💻

Code / Technical

Highly structured (to be covered later)

🎭

Simulation / Role Play

Attainable request to run a model through a simulation as a user/customer, or as a stakeholder poking holes and raising possible objections to consider.

🎯

Strategy / Decision Support

Pro/con risk mapping, prioritization with high context prompts (best to use a voice-to-text tool like Wispr)

⚙️

Task Automation / Workflow

(agents, APIs, etc. will cover later)

Structure and Sample Prompts To Try

Analysis Prompt Structure

🏆

Recommended Models:

🥇 Primary: Claude (Opus 4.1 is best, Sonnet 4 is good)

🥈 Secondary: ChatGPT ("o1-pro" is best, or GPT5 is good)

Example Uses:

  • A site for LLM SEO optimization
  • A competitor site for strategy/positioning
  • A board deck for key insights
  • A landing page for appeal and hook -- stickiness & interactivity

Structure using SEO example:

You are...

An SEO expert

I will provide...

The URL and name of the company

Your job is to...

Provide detailed feedback on site SEO opportunities with an eye for additional LLM traction based on product or company benefits that may not be clearly spelled out.

PHASE 1 - discovery

  • Search/review (define sources)
  • Pull (filter) items only, mark others as (ignore or archive)
  • Log: (pieces of information relevant) 1--2 key snippets

PHASE 2 - signals

Extract signal:

  • Unique identifiers
  • Red flags/negative sentiment
  • Current keyword focus areas

PHASE 3 - insights

Structure -- 5 sections with content and 3 to 5 bullets each

  1. Overview of SEO vs LLM optimization key differences and overarching opportunities
  2. e.g., content updates
  3. e.g., technical
  4. e.g., IA
  5. Watch Outs -- areas where it would be risky or easy to spend a lot of time for less return

PHASE 4 - receipts

Short table:

Claim | Snippet | Source (annotation) | Confidence (0--100)

RULES

  • Always include dates + sources
  • Be direct: highlight both value and risk
  • Tone: practical, insider, never salesy

OUTPUT FORMAT

  • Title: [Custom Title Here: Topic & Month Year]
  • Output in clean markdown
🎯

Try this Analysis Prompt, fill in the URL and COMPANY NAME

You are...

A senior strategist and digital analyst with a specialty in eCommerce, conversion optimization, and AI-augmented competitive intelligence.

I will provide: [The URL] and [name of the company] under review.

Your job is to...

  • Provide a competitive strategy analysis of the submitted company and top 3 identified competitors.
  • Your analysis should include clear observations on site structure, conversion strategy, brand positioning, and comparative strengths/weaknesses.

PHASE 1 -- Discovery

  • Search/review sources: brand website (all public-facing pages), recent press releases, product pages, cart/checkout flow, core social media profiles, Google snippet/SEO meta preview. Content gathered from anywhere other than brand site should be <90 days old. (archive) any older content
  • Filter strategy: Pull only relevant data tied to strategy, pricing, messaging, UX, and product structure.
  • Tag extra info as: (ignore) if not relevant to the prompt or (archive) if worth saving but not immediately useful.
  • Log: 1-2 sentences for each major insight or snippet - If visual content is referenced, note it as [screenshot].

PHASE 2 -- Signal Extraction

Distill strategic markers, flags, and intent signals.

  • CRO Signals - Notable wins (e.g., urgency cues, tiered pricing, free shipping thresholds, reviews UX), friction points or gaps.
  • OOS Flags - Identify any out-of-stock issues on core SKUs or variant SKUs; estimate impact on perceived availability/scarcity.
  • Keyword Focus - Signs of SEO or paid keyword targeting; is it aligned with category/brand goals?
  • Marketing Position - Brand voice, visual consistency, value prop clarity; evaluate across PDPs, homepage, and social.
  • Price Tiering - Identify perceived tier (luxury/mid/value), average price range, estimated AOV.
  • Discounting Strategy - Note presence of discount messaging, default sale price vs original price, email pop-ups, bundling.

PHASE 3 -- Competitive Insights

Structure your core findings in five sections, each with 3--5 bullets.

  1. Positioning & Messaging
    • Primary value props
    • Visual/voice coherence
    • Differentiation vs category norms
  2. CRO & UX Highlights
    • What works well
    • Where friction exists
    • Mobile vs desktop consistency
  3. Pricing & Promotions
    • Price tier & avg SKU price
    • Suspected AOV
    • Discount behavior (always-on vs tactical)
  4. Competitor Contrast (Top 3) List 3 competitors and include:
    • One-line description
    • 3 bullets each: key differentiators
  5. Watch-Outs
    • Areas of overinvestment or unclear ROI
    • UX or trust issues
    • Strategic misalignment to audience or category

PHASE 4 - Receipts

Construct a short table:

Claim | Snippet | Source (annotated) | Confidence (0--100)

Mark any older or secondary info as (archive) where appropriate.

Rules:

  • Always include dates and URLs when possible
  • Tone should be blunt, practical, and never salesy
  • Output format is clean markdown
  • Leave the Company and URL fields blank for user input
  • Title the output: [Company Name] - Competitive Strategy & Positioning Analysis -- [Month Year]

Image Gen + Video uses and Structure

LLM Model Notes:

  • ChatGPT as Primary/Free & Easy (GPT5/Sora)
  • Midjourney is best option for fantastical/surreal/art (there are others)
  • Gemini is best option for ultra realistic
  • Gemini VEO3 or 3.1 -- good for ad content or video

Example Uses:

  • Ad/site creative -- Gemini with Silo/photography for image gen and video
  • Fun images/art -- GPT/Midjourney
  • Backgrounds to place product etc. any/trial

Image prompt structure:

[Subject & Action], [Scene/Setting], [Lighting/Time], [Style Keywords], [Camera Type/Angle], [Color Palette], [Details] --ar [aspect ratio use 4:5 for social posts] --style [raw/expressive] --v [version -- if no input will use most recent] --p [use if you have a custom profile set up]

General Notes:

  • The Version tag is unique to Midjourney but will work on other models
  • Use most important words first
  • Avoid commas if you want the image gen to interpret loosely
  • If using a reference image in a URL, put the URL first
🎨

Try this Image Gen Prompt

Midjourney is IDEAL: https://www.midjourney.com

Can use ChatGPT: https://chatgpt.com/

A wide, cinematic digital painting of calm blue-gray water with kraken tenticles reaching up out of the water in the distance. There are towering cliffs covered in overgrowth and vines in the background, a single ruin---distorted at the edges---overlooks the sea. No pines. The atmosphere is overcast with some evening sun hitting the cliffs. highly detailed, fine digital art. raw, fantasy realism. --ar 4:5 --v 7

Both Midjourney and Gemini are IDEAL: https://www.midjourney.com https://gemini.google.com/app

Photograph version:

13w frontal real-life photograph, ultra-high-resolution 16k, Canon photography style. A soft, airy scene with bright, neutral beige background wall. A single small decorative watercolor painting of a strawberry (27x35cm), framed in matte black wood with glass reflection, centered on the wall. Linen canvas texture. Soft late afternoon sunlight shines from the left, casting a gentle shadow. Below the painting: a wide, dark gray vintage French-style wooden drawer cabinet. On the cabinet: a stack of books (left), a textured blue ceramic vase (right). Minimalist, high-fidelity texture, serene lighting. --ar 16:9 --style raw

Illustration version:

Airy and bright interior illustration in a soft, pastel palette. A neutral beige wall with a small framed watercolor painting of a strawberry (27x35cm), black wooden frame, subtle glass reflection. Linen-textured canvas. Late afternoon light filters from the left, casting soft shadows. Below: wide, dark gray French vintage drawer cabinet. A neat stack of books sits on the left; on the right, a blue ceramic vase with a textured finish. Light, peaceful atmosphere, delicate detail, soft shadows. --ar 16:9

Video Prompts

Video Prompt Structure:

Prompt: [Action or motion], [Subject], [Scene/Environment], [Time of day or lighting], [Mood or tone], [Cinematic/Style terms], [Camera movement or type]

Reference or upload image: [uploaded or URL] (optional)

Parameters: duration, aspect ratio, style (cinematic, hyperreal, animation), camera (tracking shot, dolly zoom, etc.)

🎬

Try this Video Gen Prompt

Use Gemini https://gemini.google.com/app to create video: create an account or sign in, and make sure you have video selected in bottom of the prompt box

A woman in a flowing red dress walks confidently through a crowded city street filled with people in identical black suits, all moving in the opposite direction. She moves in slow motion, unfazed, as heads subtly turn to watch her. Midday overcast lighting casts soft shadows. Inspired by the iconic red dress scene from The Matrix. Hyperreal, cinematic tone. Camera follows her in a fluid tracking shot from behind, then swings to a side angle revealing faces in passing. Crisp depth of field, sharp contrast between red and monochrome surroundings.

Settings: 10s duration, 16:9, hyperreal style