Principles are abstract until you see them applied. This guide walks through concrete before-and-after prompts across real tasks, with a breakdown of what made each one work or fail. You will see the weak version, the strong version, and the specific changes that closed the gap.
We picked tasks people actually do: summarizing documents, drafting emails, extracting structured data, brainstorming, and rewriting. For each, we show the lazy prompt most people start with, then the engineered version, and explain the reasoning. The point is not to memorize these prompts but to internalize the moves so you can apply them to your own work.
If you have read our best practices guide, this is where those practices come to life. If you have not, you can still follow along; each example stands on its own.
Example 1: Summarizing a Long Document
Weak prompt: "Summarize this article. [paste 2,000 words]"
Strong prompt: "Summarize the article between the tags for a busy executive who will not read the original. Give one paragraph under 120 words covering the main argument, then three bullets with the most decision-relevant facts. Use only information from the text. <article>...</article>"
What changed and why
We added an audience (busy executive), a length cap, a specific structure, and a grounding constraint. The weak prompt gets an undifferentiated wall of text; the strong one gets something an executive can act on in thirty seconds. The delimiter prevents the model from treating the article as instructions.
Example 2: Drafting a Client Email
Weak prompt: "Write an email telling the client the project is delayed."
Strong prompt: "Write a short, professional email to a client named Marcus telling him our redesign project will slip two weeks because of a vendor delay on our end. Apologize once without groveling, take responsibility, give the new delivery date of October 14, and offer a 20-minute call this week. Warm but concise, under 150 words."
The weak version produces a generic template you will heavily edit. The strong version is usable almost as-is because it carries the actual facts, the relationship tone, and the desired action. Notice the tone guidance ("apologize once without groveling") that steers away from the over-apologetic default. Our how-to guide breaks down this kind of prompt construction step by step.
Example 3: Extracting Structured Data
Weak prompt: "Pull out the key info from these customer reviews."
Strong prompt: "From the reviews between the tags, extract one row per review in this exact format: Sentiment (positive/negative/mixed) | Main complaint or praise | Product feature mentioned. If a field is not present, write 'none'. Do not infer feelings not stated. <reviews>...</reviews>"
This is where showing the format pays off. The weak prompt yields prose you cannot paste into a spreadsheet. The strong prompt yields a clean table because it specifies the exact columns, the allowed values, and what to do with missing data. The "do not infer" line prevents the model from hallucinating sentiment the reviewer never expressed. The common mistakes guide explains why describing format fails where showing it succeeds.
Example 4: Brainstorming With Constraints
Weak prompt: "Give me some marketing ideas for my coffee shop."
Strong prompt: "Suggest eight low-budget marketing ideas for an independent coffee shop in a college town, each costing under $200 to launch. For each, give a one-line description and the single biggest risk. Avoid generic ideas like 'use social media'; be specific and slightly unconventional."
Unconstrained brainstorming returns the obvious ideas everyone already knows. The constraints, budget cap, specific setting, required risk note, and the explicit ban on generic suggestions, force the model toward ideas worth having. Asking for the risk alongside each idea makes the output decision-ready rather than just a list.
Example 5: Rewriting for Tone and Clarity
Weak prompt: "Make this sound better. [paste paragraph]"
Strong prompt: "Rewrite the paragraph between the tags to be clearer and more direct. Cut jargon, use shorter sentences, and keep it under the original length. Preserve all facts and the professional tone. Return only the rewrite, no commentary. <text>...</text>"
"Make this better" is the vaguest possible instruction and produces unpredictable changes. The strong version defines what "better" means here (clearer, shorter, jargon-free), sets guardrails (preserve facts and tone), and suppresses the model's habit of adding "Here is your improved version!" preamble.
Example 6: Turning Notes Into a Decision Brief
Weak prompt: "Help me figure out which vendor to pick from these notes."
Strong prompt: "Using only the notes between the tags, produce a decision brief. First list the evaluation criteria mentioned, then score each of the two vendors against each criterion as strong, adequate, or weak, then give a one-paragraph recommendation with the single biggest risk of that choice. Do not introduce criteria or facts not in the notes. <notes>...</notes>"
This example combines several moves at once. It forces reasoning before the conclusion by requiring criteria and scoring first, which produces a more defensible recommendation than asking for the answer cold. The grounding constraint keeps the model from importing generic vendor-selection wisdom that may not apply. The required risk note turns a recommendation into something you can actually weigh. This is the kind of structured, reasoning-first prompt our framework guide formalizes.
What These Examples Have in Common
Look across all five and the same moves repeat:
- A specific audience or use context instead of a general one.
- An exact output format, often shown rather than described.
- A grounding or preservation constraint to prevent fabrication or drift.
- A ban on the obvious default, whether generic ideas or chatty preamble.
These four moves account for most of the gap between an amateur prompt and a professional one. Apply them to any task and your output quality jumps immediately. The framework guide organizes these moves into a repeatable model.
Frequently Asked Questions
Can I just copy these prompts for my own work?
You can adapt them, but the real value is learning the moves: add audience, specify format, add a grounding constraint, ban the default. Once you internalize those, you can engineer a strong prompt for any task rather than searching for a pre-made one that matches your exact situation.
Why do you keep wrapping content in tags?
Tags separate the data you want processed from the instructions about how to process it. Without that separation, the model can confuse a pasted document for a set of commands, which produces strange results. The tags are a small habit that prevents a frustrating class of failures.
How do I stop the model from adding "Here is your answer" preamble?
Add an explicit instruction like "Return only the rewrite, no commentary" or "Output the table and nothing else." The chatty preamble is a default behavior, and a direct instruction to suppress it works reliably across most tasks.
What makes the brainstorming example work better than a generic request?
The constraints. A budget cap, a specific setting, a required risk note, and an explicit ban on obvious ideas all push the model away from the average answer toward something useful. Unconstrained brainstorming returns the ideas you already thought of yourself.
Do these examples work for any AI model?
The structure transfers across models because it reflects how the request is framed, not a trick specific to one tool. You may need to adjust small details, but adding audience, format, and constraints improves output regardless of which model you use.
Key Takeaways
- Weak prompts share one trait: they leave audience, format, and constraints to chance.
- Add a specific audience and use context to escape generic output.
- Show the exact output format, especially for structured data extraction.
- Include grounding or preservation constraints to prevent fabrication and drift.
- Explicitly ban the default, whether obvious ideas or unwanted preamble.