When the Curiosity rover landed on Mars on August 6, 2012, NASA engineers at the Jet Propulsion Laboratory had spent years writing extraordinarily specific instructions. Every command sent to the rover specified exact coordinates, force measurements in newtons, timing windows to the millisecond. Nothing was left to interpretation. The rover could not ask a follow-up question. Vague instructions would have meant a $2.5 billion spacecraft sitting still — or worse.
AI language models face a similar problem in reverse: they receive your words and must infer everything you left unsaid. Unlike NASA engineers, most people never learn to write with that level of precision — until now.
Every prompt sits somewhere on a spectrum from completely vague to richly specific. The position on that spectrum directly determines how useful the AI's response will be. This is not a matter of opinion — it is a measurable, reproducible effect.
In 2023, researchers at Stanford's Human-Centered AI group published findings showing that adding just three specific constraints to a prompt — audience, format, and goal — increased user satisfaction with AI responses by over 60%. The words themselves mattered less than whether they carried concrete information.
AI models are trained on billions of text samples. When you send a vague prompt, the model must choose which of thousands of possible interpretations to follow. It defaults to the statistical average — the most common answer to the most common version of your question. That average is rarely what you specifically need.
Think of it like a treasure map with no landmarks. "Go north and find the chest" could apply to every patch of ground on the planet. Add "walk exactly 40 paces north from the old oak at the park entrance" and suddenly you are digging in one specific spot. Specificity in prompts works the same way — it narrows the search space from infinite to useful.
In 2001, Google's engineers noticed that one-word searches returned terrible results. Users who typed "jaguar" wanted either the car or the animal — the engine couldn't tell. By 2024, Google processes queries averaging 4.7 words, and internal satisfaction data shows longer, more specific queries produce dramatically better outcomes. The same principle governs AI chat: more context, better results.
In this module, we treat specificity as a treasure hunt. Each lesson gives you a new "clue ingredient" — a specific type of detail that makes your prompts more powerful. By the end, you will have a complete map: a set of ingredients you can add to any prompt to unlock the exact response you need.
The game has four rounds: understanding what vague looks like, learning the five specificity ingredients, practicing the before/after rewrite, and finally running live tests to see the difference yourself.
Specificity is not about using more words. It is about including the right information — the details that answer the questions the AI would ask if it could ask them. Every vague word in your prompt is an unanswered question that the AI must guess at.
This lab is about seeing the problem firsthand. Start by sending a deliberately vague prompt, then send a more specific version and compare the responses. The AI assistant will help you analyze the difference.
Try at least 3 exchanges. Suggested starting point:
When OpenAI prepared GPT-4 for public release in March 2023, internal red-team testers discovered a recurring pattern: the most problematic outputs — both too harmful and too useless — came from prompts that were missing the same categories of information. The testers began tagging prompts by what they lacked. The most common gaps were who the audience was, what format was needed, what the actual goal was, how long the response should be, and what tone was appropriate.
These five gaps became an informal checklist inside the team. External prompt engineering courses — including those from Anthropic, Google DeepMind, and Cohere — later codified similar frameworks. The specificity ingredients below are drawn directly from those published guides.
Think of writing a prompt like cooking a recipe. A dish with one ingredient is edible but boring. Add five well-chosen ingredients and you get something memorable. Here are the five specificity ingredients:
Who will read or use the response? Age, background, expertise level, and emotional state all matter.
What shape should the answer take? Bullet list, essay, table, step-by-step instructions, dialogue, code?
What will the response be used for? To teach, persuade, entertain, summarize, compare, or create something?
How long should the response be? A single sentence, a paragraph, 200 words, a full page?
What emotional register? Formal, casual, encouraging, urgent, neutral, playful, serious, empathetic?
You do not always need all five. A quick factual question might only need a length constraint. A sensitive email might prioritize tone and audience. The skill is knowing which ingredients your specific prompt is missing — that is where the quality gap lives.
In 2022, Anthropic published a guide for using Claude (their AI) that noted prompts with at least three of these five ingredients produced measurably more "on-target" responses — ones users accepted without editing. Prompts with zero or one ingredient led to users refining the response three or more times on average before it was useful.
Before: "Write me an email about the project delay."
After: "Write a 3-paragraph professional email to my project stakeholders (senior managers who are not technical) explaining that the software launch is delayed by two weeks due to a third-party integration issue — tone should be calm and solution-focused, not apologetic."
The "after" version has all five ingredients. The "before" has none.
Think of each of the five ingredients as a clue in the treasure hunt. Every clue you add moves you closer to the buried chest — the exact response you need. A prompt with zero clues leaves you wandering. A prompt with all five sends you directly to the X.
A = Audience · F = Format · G = Goal · L = Length · T = Tone
Memory trick: "A Fine Goal Lasts Today" — Audience, Format, Goal, Length, Tone.
Practice diagnosing prompts. Send a prompt to the AI and ask it to identify which of the five ingredients (Audience, Format, Goal, Length, Tone) are present and which are missing. Then rewrite the prompt to include all five.
Try at least 3 exchanges. Start with one of these examples or use your own:
When GitHub released Copilot in June 2021 — an AI coding assistant built on OpenAI's Codex — early users struggled to get useful code suggestions. A developer survey conducted by GitHub in early 2022 found that users who got the best results had developed a personal habit: they never sent their first instinct. Instead, they wrote a rough prompt, read it back, added context about the programming language, the function's purpose, the expected input/output format, and any edge cases to handle — then sent the revised version.
GitHub published these patterns in a guide called "How to write better prompts for GitHub Copilot" in September 2022. The core technique was the same rewrite loop: draft, identify what's missing, add it, send. That three-step loop is what this lesson teaches.
Professional prompt engineers rarely get a great response on the first try — and they know it. The difference is that they have a reliable rewrite process rather than just trying again with slightly different words and hoping for the best.
The loop has three steps:
| Before (Draft) | What AI Must Guess | After (Rewritten) |
|---|---|---|
| "Write something about healthy eating." | For whom? What format? How long? What aspect of healthy eating? What will it be used for? | "Write a 5-bullet-point cheat sheet on reducing sugar intake, aimed at busy parents with no nutrition background, for printing and sticking on a fridge." |
| "Explain machine learning to me." | What do I already know? How deep? What's my goal — casual curiosity or professional use? How long? | "Explain what machine learning is in 3 sentences for someone who knows Excel but has never studied programming. Goal: I'm deciding whether to take an online ML course." |
| "Make this email better." | 'Better' how? More formal? Shorter? More persuasive? Who is receiving it? | "Revise this email to be more concise (under 100 words) and more persuasive. The recipient is a skeptical VP who values data. Keep a professional tone." |
The most powerful question you can ask about any prompt you write is: "What would the AI have to guess to answer this?" Every guess is a gap. Every gap is a place where the response might miss your actual need.
In 2023, Ethan Mollick — a Wharton professor who studies AI adoption — published research showing that workers who explicitly tested their prompts for hidden assumptions before sending them got better outcomes than those who iteratively refined responses after the fact. The upfront investment in a rewrite saved time overall.
You now have two clues: the five specificity ingredients, and the three-step rewrite loop. Together, they give you a method — not just a list of things to add, but a process for finding what's missing. The next lesson turns this into a live game you play against your own prompts.
Adding length without adding specificity. "Write something about healthy eating for busy parents who are interested in reducing their sugar intake and want to learn more about nutrition in a way that is easy to understand" — this is longer but still vague. Length is not specificity. Concrete details are.
Rewriting tone when you need format. If the AI's response is well-written but structured wrong (an essay when you needed bullet points), tone editing won't help. Identify which ingredient is actually missing.
Adding context the AI doesn't need. Backstory and emotional context are sometimes useful for tone calibration, but they shouldn't replace the core five ingredients. Keep rewrites focused.
Use the three-step rewrite loop on real prompts. Send a before version, then ask the AI to help you identify what it had to guess, then send the rewritten version and compare results.
Try at least 3 exchanges. Start here:
In 2023, NHS England piloted AI-assisted clinical documentation tools across several hospital trusts. A recurring problem emerged in the first months: clinicians produced wildly inconsistent AI outputs from the same underlying tools. Investigation revealed that doctors who got useful summaries had learned — often informally from each other — to include patient age bracket, the document's purpose (referral vs. internal note), and the required length in their prompts. Those who hadn't learned these habits sent one-line prompts and got one-size-fits-all responses that required heavy editing.
NHS Digital subsequently issued informal guidance recommending that clinical staff treat AI prompting like filling in a patient intake form — every field exists for a reason, and leaving blanks creates downstream problems. The specificity framework was, in essence, their solution.
You now have all four clues collected across this module. Here is the complete map:
Vague prompts force AI to guess. Every guess is a gap between what you need and what you get.
Ask yourself: "Is my prompt specific enough to have only one reasonable interpretation?"The five ingredients fill the most common gaps: Audience, Format, Goal, Length, Tone.
Ask yourself: "Which of the five ingredients am I missing?"The three-step rewrite loop (draft → identify guesses → replace with facts) is a repeatable process.
Ask yourself: "What would the AI have to guess if I sent this now?"Specificity is a habit, not a one-time fix. Apply it to every important prompt before sending.
Ask yourself: "Is this the first thing I thought of, or is it what the AI actually needs?"Not every prompt needs all five ingredients. The key is knowing when to invoke each one:
Audience matters most when the complexity or vocabulary of the response needs to match a specific person. A response for a child and an expert on the same topic should be completely different.
Format matters most when you will use the output in a specific way — a report, a slide deck, a code file, a social post. If format is open, you can leave it out.
Goal matters most for persuasive, creative, or strategic tasks where the AI's output needs to serve a larger purpose beyond the prompt itself.
Length matters most when you have a constraint — a word limit, a time limit, a character count, a conversation budget.
Tone matters most for human-facing communication — emails, speeches, marketing copy, care messages — where emotional register affects the reader's response.
Google's internal testing of Bard (now Gemini) before its 2023 public launch found that test users who added even one well-chosen specificity ingredient to their first-draft prompt were three times less likely to abandon a conversation mid-task. The single most impactful ingredient was Format — telling the AI what shape the answer should take reduced confusion about whether the response was useful.
The treasure you are hunting in this module is not a perfect prompt template. It is a flexible mental checklist — four questions you ask yourself before sending any important prompt:
1. Could this be interpreted two different ways? · 2. Which of A-F-G-L-T am I missing? · 3. What would the AI have to guess? · 4. Have I replaced the guesses with facts?
That checklist is the X. Every skilled AI user in every field — from NHS clinicians to GitHub developers to Wharton researchers — has arrived at some version of it. Now you have it too.
Specificity is the single highest-leverage skill in prompt writing. It costs no extra technology, no special access, and no subscription upgrade. It costs only the habit of pausing before you send, and asking four questions. The AI you already have access to will produce dramatically better results the moment you start using it.
Run the full process end-to-end. Pick a real task you actually want help with — anything. Write a first-instinct prompt, then apply the four-question checklist, rewrite it with all five ingredients, and send both versions. The AI will help you analyze the difference in results.
Try at least 3 exchanges. Start here: