L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Module 2 · Lesson 1

Why Words Are Everything

The quality gap between vague and precise prompts is larger than most people ever discover.
What actually separates a prompt that works from one that wastes your time?

When the GitHub Copilot research team published their findings in August 2022, they documented something striking: developers who wrote descriptive function names and clear inline comments before invoking AI completions accepted usable code at nearly twice the rate of developers who used terse, vague identifiers. The AI hadn't changed. The prompts had.

The Specificity Gap

Every AI model — whether it's generating text, images, or code — is performing a form of pattern completion. It fills in the most statistically likely continuation of what you've given it. When you give it almost nothing, it reaches for the most generic, averaged-out response it can construct.

This is why "write me a cover letter" produces something that feels like it was assembled from a recycling bin of HR clichés, while "write a cover letter for a mid-career data analyst applying to a Series B fintech startup, emphasizing their work reducing reporting latency by 40%" produces something usable on the first draft.

The difference isn't luck. It's specificity. Specificity constrains the solution space, steering the model away from the averaged-out middle and toward the precise corner of response-space you actually need.

Weak Prompt
"Write something about climate change for my blog."
Strong Prompt
"Write a 400-word blog intro for a sustainability-focused small business audience explaining why the SEC's new climate disclosure rules matter to them. Conversational tone, no jargon."
The Three Dimensions of Weak Prompts

Weak prompts fail along three predictable axes. Understanding them helps you diagnose your own prompts before you send them.

Underspecified Task The model doesn't know what "done" looks like. "Help me with my email" could mean proofread, rewrite, summarize, or respond — it guesses.
Missing Audience Without knowing who the output is for, the model defaults to a generic professional register that fits nobody in particular.
No Format Signal Length, structure, tone, and formality are all left to chance. You get whatever the training distribution considered average.
Real Pattern — OpenAI's Usage Research, 2023

OpenAI's published analysis of ChatGPT usage patterns found that users who provided context about their role, goal, and desired format in their initial prompt required significantly fewer follow-up clarifications before reaching a usable output — reducing the average conversation length to reach a satisfactory result by roughly half compared to bare-minimum prompts.

What Specificity Actually Costs You

The counterintuitive truth is that writing a more specific prompt takes perhaps thirty extra seconds, but saves you multiple rounds of regeneration, editing, and frustration. The economics strongly favor investment upfront.

Think of it this way: a painter who tells their assistant "get me a brush" will spend more time running back to the supply room than one who says "get me the 1-inch flat synthetic bristle brush." The specificity cost is seconds. The vagueness cost is minutes — and a wrong output you now have to correct.

This module is about training that thirty-second habit. Every lesson builds one component of a high-quality prompt until, by the end, precise prompting feels as natural as typing.

Core Principle

A prompt is a specification. The more precisely you specify what you want, the less the model has to guess — and the less it guesses, the less it surprises you with something you didn't ask for.

In Lesson 2, we'll break apart the anatomy of a complete prompt and assign a function to each component — so you can diagnose and repair your own prompts systematically.

Module 2 · Lesson 1

Quick Check — Why Words Are Everything

3 questions · select the best answer for each
1. According to the GitHub Copilot research cited in the lesson, what doubled acceptance rates for AI-generated code?
Correct. The study found that contextual clarity in the code itself — before any explicit prompt — led to nearly twice the rate of usable completions.
Not quite. The key finding was that descriptive naming and comments gave the model clearer context, dramatically improving its output quality.
2. What does an AI model do when given a vague prompt with almost no context?
Correct. With little to constrain it, the model fills in the most statistically average response — producing generic output that fits no specific need well.
Not quite. The model doesn't refuse or randomize — it produces what's most statistically likely given the minimal input, which is usually a generic, averaged result.
3. Which of the following is identified as one of the three dimensions of weak prompts?
Correct. Missing audience, underspecified task, and no format signal are the three core failure modes of weak prompts identified in this lesson.
Not quite. The three dimensions are: underspecified task, missing audience, and no format signal. Length and technical vocabulary aren't inherently weaknesses.
Module 2 · Lab 1

Diagnosing Weak Prompts

Practice identifying and fixing the three failure modes

Your Lab Task

You'll practice diagnosing weak prompts and rewriting them to be specific, audience-aware, and format-clear. Share a weak prompt you've used before (or make one up), and your lab assistant will help you identify exactly what's missing and how to fix it.

Complete at least 3 exchanges to mark this lab done.

Start by sharing a vague or weak prompt you've tried — something that gave you a mediocre result. Or just type: "Give me a weak prompt to diagnose."
Prompt Diagnosis Lab L1
Welcome to the Prompt Diagnosis Lab. Share a weak or vague prompt you've actually used — or one you make up — and I'll walk you through exactly what's missing and how to repair it. What have you got?
Module 2 · Lesson 2

Anatomy of a Complete Prompt

Every strong prompt has identifiable parts. Learn to build each one deliberately.
What are the five components that transform a bare request into a precise specification?

When Midjourney's Discord communities began sharing prompting guides in late 2022, the most shared documents weren't about creativity — they were taxonomies. Users had reverse-engineered that the AI responded to at least five distinct types of information: the subject, the style, the medium, the mood, and the technical parameters. Prompts that addressed all five produced images users described as "exactly what I had in mind." Prompts that addressed one or two produced something — just not that.

The Five-Part Prompt Framework

This framework applies across AI tools — text generators, image tools, code assistants, and voice interfaces. Each component answers a question the model would otherwise have to guess.

Complete Prompt Anatomy
Role Who should the AI be? "You are a senior UX writer at a B2B SaaS company." Role primes the vocabulary, register, and expertise level of the response.
Task What specific action should it perform? "Write the empty-state copy for a dashboard with no data yet." Not "help with copy" — a specific, bounded action.
Context What does it need to know to do this well? "Our users are non-technical operations managers who feel anxious when software feels broken." Background that changes how the task should be approached.
Format What should the output look like? "Give me a headline (6 words max), one sentence of explanation, and a single CTA button label." Length, structure, components.
Constraint What should it avoid or stay within? "Don't use the word 'yet'. Keep it warm, not cutesy." Explicit guardrails prevent the most common failure modes.
Not Every Prompt Needs All Five

The framework isn't a checklist you must complete before every interaction. It's a diagnostic tool. When you get a bad output, ask: which component is missing?

A simple factual question — "What year did the Berlin Wall fall?" — needs none of them. But a request for creative, professional, or technical output almost always benefits from at least three: task, context, and format. Role and constraint are the power-ups you add when stakes are higher.

Documented Example — Anthropic's Prompt Library, 2024

Anthropic's publicly released prompt library for Claude — used by developers building on their API — consistently uses all five components for complex tasks. Their "Meeting Summarizer" template specifies role (expert meeting analyst), task (extract key decisions and action items), context (the meeting transcript and its purpose), format (structured bullets with owner names), and constraint (exclude small talk and pleasantries). The template pattern proved reliable enough across thousands of API calls that it became their recommended starting structure for new developers.

Building a Prompt Live

Watch how each component transforms a bare request into something the model can act on precisely:

Bare Request
"Write an email to a client."
+ Task
+ Task
"Write an email to a client explaining that their project will be two weeks late."
+ Context
+ Context
"…The delay is due to a supplier issue outside our control. The client is a long-term account we value highly."
+ Format
+ Format
"…Keep it under 150 words. Three paragraphs: acknowledgment, reason, revised timeline and next step."
+ Constraint
+ Constraint
"…Don't be defensive or over-apologetic. Professional but warm tone. No corporate jargon."
The Payoff

Each component you add narrows the solution space. The final prompt leaves the model with almost no room to surprise you — it has to produce approximately what you asked for. That's the goal.

In Lesson 3, we'll zoom in on role-setting — the single highest-leverage component for changing the quality register of an AI's output — and see documented examples of how it shifts results dramatically.

Module 2 · Lesson 2

Quick Check — Anatomy of a Complete Prompt

3 questions · select the best answer for each
1. In the five-part prompt framework, what does the "Context" component provide?
Correct. Context gives the model background that shapes its approach — audience details, situational facts, stakes — not what to do, but what to know while doing it.
Not quite. That describes other components. Context is the background knowledge the model needs to do the task well — facts that change how it should approach the request.
2. What is the primary purpose of including a "Role" in your prompt?
Correct. Assigning a role tells the model what kind of expert voice to adopt, which fundamentally shifts the vocabulary, depth, and register it uses.
Not quite. Role is about setting the expert persona — the vocabulary, depth, and register. Safety, format, and background facts are handled by other components.
3. According to the lesson, when does a prompt NOT need all five components?
Correct. A simple factual question like "What year did the Berlin Wall fall?" needs none of the five components. The framework is a diagnostic tool, not a mandatory checklist.
Not quite. The lesson is clear that the framework is a diagnostic tool — not a mandatory checklist. Simple factual questions don't need it; complex outputs do.
Module 2 · Lab 2

Building Complete Prompts

Practice assembling all five components from scratch

Your Lab Task

Pick any real task you've wanted AI help with — at work, in school, or for a personal project. Your lab assistant will guide you through adding each component (Role, Task, Context, Format, Constraint) to build a complete, high-quality prompt.

Complete at least 3 exchanges to mark this lab done.

Tell me a task you'd like AI help with, and we'll build a complete five-component prompt together — one layer at a time.
Prompt Builder Lab L2
Welcome to the Prompt Builder Lab. Tell me about a task you actually want help with — writing, coding, research, design, anything — and I'll guide you through adding each of the five components until you have a complete, precise prompt. What task are you working with?
Module 2 · Lesson 3

Role-Setting and Tone Control

The single fastest way to shift the quality of an AI output is to tell it who to be.
How does assigning a role change what the AI produces — and what makes a good role definition?

In a 2023 study at Stanford's Human-Centered AI Institute, researchers tested how role-priming affected the perceived quality of GPT-4 outputs across medical, legal, and financial domains. Outputs prefaced with a specific expert role were rated by domain professionals as significantly more appropriate in tone and vocabulary than outputs without role-setting — even when the factual content was identical. The AI didn't know more. It just sounded like someone who did.

Why Roles Work

Large language models are trained on vast amounts of text written by people with many different roles, expertise levels, and communication styles. When you assign a role, you're essentially activating a subset of that training data — filtering the model toward patterns associated with that type of writer or expert.

A prompt beginning with "You are a pediatric nurse explaining this to an anxious parent" will produce substantially different language than "You are a medical researcher summarizing this for a journal abstract" — even if both are asked about the same condition. The facts can be the same. The framing, vocabulary, and level of assumed knowledge will differ completely.

No Role — Generic
"Explain how compound interest works."

→ Produces a textbook-style paragraph that's accurate but dry, pitched at no one in particular.
With Role — Targeted
"You are a high school personal finance teacher. Explain compound interest to a 16-year-old using a relatable example about saving for a car."

→ Produces an engaging, age-appropriate explanation with narrative momentum.
What Makes a Strong Role Definition

A strong role definition has at least two components: the position or expertise type and the context or audience relationship. The position activates the vocabulary. The context activates the communication style.

Weak Role "You are an expert." — Too vague. Expert in what? For whom? This barely constrains anything.
Medium Role "You are a marketing expert." — Better, but still broad. Marketing for B2B SaaS differs from marketing for consumer packaged goods.
Strong Role "You are a senior content strategist at a B2B SaaS company, writing for a non-technical VP of Operations audience." — Position + domain + audience relationship.
Documented Case — Bing Chat Enterprise Prompting Guide, 2023

Microsoft's internal prompting guide for enterprise Bing Chat users — shared publicly in their documentation — explicitly advised users to open prompts with role definitions for any professional task. Their documented example compared "summarize this document" with "You are a business analyst. Summarize this document for an executive audience, highlighting decisions required." Internal testing showed the role-primed version consistently produced summaries rated as "immediately actionable" by business users.

Tone as a Separate Lever

Role and tone overlap but aren't identical. A role sets who the AI is. Tone sets how that person communicates. You can have the same role with different tones — a "senior software engineer explaining this to a new hire" can be warm and encouraging, or direct and technical, or even a bit Socratic (asking questions rather than giving answers).

When stakes are high, specify both:

Role + Tone Example
Role You are a senior product manager at a consumer tech company.
Context You're writing to a new engineering team that has strong technical skills but limited experience with user research.
Tone Collegial and direct. No condescension. Assume intelligence, not experience.
Practice Pattern

For any important output, try writing the role before anything else. Ask yourself: who would be the ideal human author of this piece? Then describe that person in two to three specific attributes. Position + domain + audience relationship covers most cases.

Lesson 4 puts everything together through the lens of iteration — because even a perfect first prompt rarely produces a perfect first output, and knowing how to refine efficiently is the final skill that separates one-shot users from expert prompters.

Module 2 · Lesson 3

Quick Check — Role-Setting and Tone Control

3 questions · select the best answer for each
1. What did the Stanford HAI research (2023) find about role-primed AI outputs in professional domains?
Correct. The key finding was that role-priming affected perceived quality through tone and vocabulary — not through improved factual content, which was held constant.
Not quite. The study held factual content constant. The finding was specifically about tone and vocabulary appropriateness — not accuracy, length, or human indistinguishability.
2. What are the two components that make a role definition "strong" according to the lesson?
Correct. Position activates the right vocabulary; audience relationship activates the right communication style. Together they produce a fully specified persona.
Not quite. The lesson specifies position/expertise type (which activates vocabulary) plus context/audience relationship (which activates communication style). Those two together are what make a role strong.
3. How does "tone" differ from "role" in prompt design?
Correct. The same role can be expressed in very different tones — warm vs. direct, Socratic vs. declarative. They are independent levers that can be tuned separately.
Not quite. Role and tone are distinct. Role = who the AI is (the persona and expertise). Tone = how that persona communicates (warmth, directness, formality). You can tune them independently.
Module 2 · Lab 3

Role and Tone Experiments

See how the same task changes when you vary who's answering

Your Lab Task

Pick a task — explaining a concept, writing a message, summarizing something — and try it with different role definitions. Your lab assistant will show you how the output changes, and help you identify which role + tone combination produces the best result for your actual need.

Complete at least 3 exchanges to mark this lab done.

Tell me a topic or task. I'll demonstrate how different role definitions change the output — and help you find the role + tone combination that works best for your purpose.
Role & Tone Lab L3
Welcome to the Role & Tone Lab. Give me any topic or task — something you'd actually want to explain, write, or summarize — and I'll show you how different role definitions change the output. What are we working with?
Module 2 · Lesson 4

Iteration: The Expert's Secret

Expert AI users don't write perfect prompts. They iterate faster and more deliberately than everyone else.
What separates someone who gets useful AI output on the third try from someone who gives up on the first?

Ethan Mollick, professor at the Wharton School, documented in his widely circulated 2023 research on AI in professional tasks that the highest-performing users shared a specific behavior: they treated AI interactions as dialogues, not commands. Where average users submitted one prompt and evaluated the output as pass/fail, expert users diagnosed what the first response got right, identified the specific gap, and crafted targeted follow-up prompts. Their outputs weren't better because they knew magic words. They were better because they iterated with intent.

The Three Iteration Moves

Most people who "don't get good results from AI" are actually iterating randomly — regenerating without changing anything, or abandoning after one try. Expert iterators make one of three targeted moves when an output misses:

Narrow The output was too broad or generic. Add specificity: a constraint, a concrete example, a tighter scope. "Make it shorter" or "focus only on the onboarding flow, not the whole product."
Redirect The output went in the wrong direction. Correct the assumption it made. "You assumed I was writing for a general audience — this is for experienced developers. Try again with that in mind."
Expand The output was on track but shallow. Ask it to go deeper on a specific element. "Expand on the third point — that's the one that matters most for my audience."
Diagnosing the Miss

Before you can make the right iteration move, you need to know why the output missed. The fastest diagnosis framework matches missed outputs to missing prompt components:

Output Problem → Likely Missing Component
Tone off Role or tone constraint is missing or too vague. Add a specific persona with a defined audience relationship.
Wrong thing entirely Task was underspecified. Restate exactly what action you want, what the output IS, not just what topic it covers.
Wrong assumptions Context is missing. Add the background the model assumed incorrectly, explicitly.
Bad structure or length Format was unspecified. Define length, number of items, section headers, or any structural requirement.
Included something you hate Constraint was absent. Add an explicit "do not include X" or "avoid Y."
Documented Practice — OpenAI's "Prompt Engineering Guide," 2024

OpenAI's published Prompt Engineering Guide explicitly recommends an iterative approach for complex tasks, advising users to "start with a simple prompt and iterate from there." Their documented workflow involves evaluating each output against the original goal, identifying the specific gap, and adding exactly the constraint or component that addresses that gap — rather than rewriting the entire prompt from scratch. This targeted iteration approach is described as more efficient than either single-shot prompting or complete rewrites.

Keeping What Works

One underused iteration skill is anchoring: explicitly telling the model what to keep from the previous output while changing something else. Without anchoring, a follow-up prompt can inadvertently discard good elements along with bad ones.

Compare these two follow-ups to a draft that got the structure right but the tone wrong:

Without Anchoring
"Rewrite this in a more conversational tone."

→ Model may also change the structure, examples, and length you liked.
With Anchoring
"Keep the same three-section structure and the car-payment example. Only change the tone to be more conversational — less formal, shorter sentences."

→ Model changes exactly what you specified.
The Iteration Mindset

Every output — even a bad one — gives you information. A bad output tells you which component you under-specified. A partially good output tells you what to anchor and what to change. There are no wasted prompts when you iterate with intent.

You've now completed all four lessons of Module 2. You understand why specificity matters, how to build a complete prompt from its five components, how to use role and tone as quality levers, and how to iterate with intent rather than frustration. Take the module test to confirm your understanding — then carry these skills into Module 3, where you'll apply them to real creative projects.

Module 2 · Lesson 4

Quick Check — Iteration: The Expert's Secret

3 questions · select the best answer for each
1. According to Ethan Mollick's Wharton research, what distinguished the highest-performing AI users from average users?
Correct. Mollick's key finding was behavioral — expert users diagnosed each output and made targeted follow-up moves rather than treating responses as pass/fail.
Not quite. The distinguishing factor was behavioral, not technical. Experts diagnosed each output and made targeted follow-up prompts — treating AI as a dialogue partner, not a command terminal.
2. If an AI output has the wrong tone but the correct structure and examples, which iteration move should you use?
Correct. Anchoring locks in what's working (structure, examples) while a targeted redirect changes only the tone — preventing the model from discarding good elements along with bad ones.
Not quite. When some elements are right and some are wrong, anchoring is the key technique — tell the model what to keep while specifying exactly what to change.
3. If an AI output includes things you didn't want — irrelevant sections, a cliché phrase, the wrong kind of examples — what does this most likely indicate about your prompt?
Correct. When the model includes things you dislike, the missing component is almost always a constraint — an explicit "avoid X" or "do not include Y" guardrail.
Not quite. Unwanted inclusions map to missing constraints. Adding an explicit "do not include" or "avoid" statement is the targeted fix — not removing context or simplifying the task.
Module 2 · Lab 4

Iteration Practice

Practice diagnosing outputs and making targeted iteration moves

Your Lab Task

Share a real AI output that disappointed you — or describe one from memory. Your lab assistant will help you diagnose exactly which component was missing and walk you through the targeted iteration move (narrow, redirect, expand, or anchor) that would fix it.

Complete at least 3 exchanges to mark this lab done.

Share a disappointing AI output you've received — paste the output, or describe what happened. Or just say "give me a bad output to analyze" and I'll create a scenario for us to work through together.
Iteration Diagnosis Lab L4
Welcome to the Iteration Lab. Share a disappointing AI output — something that missed the mark — and I'll help you diagnose exactly what went wrong and what targeted iteration move would fix it. Or say "give me a scenario" and I'll generate one for us to analyze together. What are we working with?
Module 2

Module Test — The Prompt Is Your Paintbrush

15 questions · 80% required to pass · covers all four lessons
1. What is the primary reason a vague prompt produces generic output?
Correct. Pattern completion with minimal constraints produces maximum-average, minimum-specific output.
Not quite. AI models don't penalize or randomly select — they produce the most statistically likely continuation, which is generically average when given little specificity.
2. The GitHub Copilot research cited in Lesson 1 found that acceptance rates for AI-generated code nearly doubled when developers did what?
Correct. Contextual specificity in the code itself — before the explicit prompt — gave the model clearer constraints to work within.
Not quite. The research specifically documented descriptive naming and inline comments as the practice that nearly doubled acceptance rates.
3. Which of the three failure modes of weak prompts describes a situation where the model doesn't know what "done" looks like?
Correct. Underspecified task means the action requested is so vague that the model has to guess what output would satisfy you.
Not quite. When the model doesn't know what "done" looks like, the failure mode is underspecified task — the action is too vague to produce a targeted result.
4. In the five-part prompt framework, what does the "Format" component specify?
Correct. Format covers structural requirements — length, number of sections or items, headers, specific components like headline + body + CTA.
Not quite. Format is specifically about structure and length — how the output should be arranged and how much of it there should be.
5. According to Anthropic's publicly released prompt library, which five-component pattern was used in their Meeting Summarizer template?
Correct. The Anthropic template used all five — role (expert meeting analyst), task (extract decisions/actions), context (transcript purpose), format (structured bullets), constraint (exclude small talk).
Not quite. The Anthropic Meeting Summarizer template documented use of all five components, which is why it became their recommended starting structure for new developers.
6. What is the recommended minimum set of components for most creative, professional, or technical output requests?
Correct. Task, context, and format cover the core gap for most complex requests. Role and constraint are the higher-stakes additions when output quality really matters.
Not quite. The lesson recommends task, context, and format as the baseline for complex outputs, with role and constraint as the power-ups for higher-stakes work.
7. What does role-setting in a prompt actually do at the model level?
Correct. Role assignment filters the model toward text patterns associated with that persona in its training data — affecting vocabulary, register, and communication style, not factual accuracy.
Not quite. Role-setting doesn't grant database access or fine-tuning. It filters the model toward patterns in its training data associated with that expert type — mainly affecting style, not facts.
8. What did the Stanford HAI (2023) research find was true about role-primed professional outputs compared to standard outputs?
Correct. The study held factual content constant — the improvement was entirely in tone and vocabulary appropriateness as rated by domain professionals.
Not quite. The Stanford research specifically controlled for factual content. The finding was that role-priming improved tone and vocabulary — not accuracy or length.
9. Which of the following is a "strong role definition" by the standard established in Lesson 3?
Correct. This includes position (senior content strategist), domain (B2B SaaS), and audience relationship (non-technical VP of Operations) — all three attributes of a strong role definition.
Not quite. A strong role needs position/expertise type AND context/audience relationship. Only the third option includes both — specifically domain and the audience being addressed.
10. What is Microsoft's documented recommendation from their Bing Chat Enterprise prompting guide about professional tasks?
Correct. Their guide explicitly advised role-primed prompts for professional tasks, documenting that role + audience produced outputs rated as "immediately actionable."
Not quite. Microsoft's documented advice was specifically to open prompts with role definitions for professional tasks — a practice that consistently improved actionability ratings.
11. According to Ethan Mollick's Wharton research, average users treated AI interactions as:
Correct. The average user submitted a prompt, evaluated it as pass/fail, and either accepted a mediocre result or gave up — while expert users iterated with intent.
Not quite. Mollick's research found average users treated AI as a command terminal with pass/fail evaluation — not as a dialogue or creative partnership.
12. Which iteration move should you use when an output was on track but too shallow — it got the right idea but didn't go deep enough?
Correct. Expand is the move when direction is right but depth is insufficient — "go deeper on point three" is more efficient than rewriting from scratch.
Not quite. When the output is directionally correct but shallow, the move is Expand — targeted deepening of the right element, not narrowing, redirecting, or restarting.
13. What is "anchoring" in the context of prompt iteration?
Correct. Anchoring prevents good elements from being inadvertently discarded when you ask for changes — you lock in what worked while targeting what didn't.
Not quite. Anchoring is specifically the technique of preserving what's working in a previous output while instructing the model to change only what's not.
14. If an AI output includes irrelevant sections or a phrase you specifically didn't want, which prompt component was most likely absent?
Correct. Unwanted inclusions map directly to missing constraints — the model will include what it thinks is relevant unless you explicitly tell it not to.
Not quite. When the model includes something you didn't want, the missing component is almost always a constraint — an explicit "avoid X" or "do not include Y" instruction.
15. What is the core principle about prompts that this module establishes?
Correct. This is the foundational principle of the module: a prompt is a specification, and specificity constrains the solution space to produce exactly what you need.
Not quite. The module's core principle is that a prompt is a specification. The more precisely you define what you want, the less the model guesses — and the better your output.