AI Prompt Engineering Crash Course

How It Works

Describe what you're trying to get an AI to do and where your current attempts are falling short. The guide maps your task to the specific patterns most likely to fix it, with a before/after rewrite you can use immediately - plus a diagnostic table so you can identify your failure mode before reading any theory.

What to Provide

Diagnose Before You Fix

Most "bad AI output" complaints trace to one of five specific, fixable causes - not a vague "the AI just isn't good enough."

The 5 Patterns That Matter Most

Replace [[tokens]] with your details. This is the finished deliverable.

Pattern 1 - Role + Goal + Constraints + Format

Before: "Write me a product description."

After: "You are a senior e-commerce copywriter. Write a product description for [[product]] targeting [[audience]]. Constraints: 100-150 words, benefit-first, no superlatives without a supporting fact. Output as a single paragraph."

Why it works: the role narrows the model's tone and vocabulary, the constraints prevent the two most common failure modes (rambling, unsupported hype), and the explicit format removes guesswork about structure.

Pattern 2 - Few-Shot Examples

Show 1-2 examples of the output style you want before asking for new work. The model matches structure and tone far more reliably from a concrete example than from a text description alone - "punchy and confident" means something different to every reader, but a sample paragraph is unambiguous.

Pattern 3 - Chain of Thought + Verify

Add this line: "Think step by step. At the end, list any assumptions you made and how I could verify them."

This surfaces the model's reasoning so you can catch a wrong assumption before it becomes a wrong answer, instead of only seeing the (possibly confidently wrong) final output.

Pattern 4 - Explicit Iteration

After the first output: "Improve the previous version: make it [[30% shorter / punchier / more technical]]. Keep [[the parts that worked]]."

Treat the first response as a draft, not a verdict. Naming exactly what to keep prevents the model from "improving" by rewriting the parts that were already correct.

Pattern 5 - Delimiters and Structure

Use triple backtick code fences or XML-style tags to clearly separate instructions from the content being transformed - this prevents the model from confusing your instructions with the input text, which is a common source of the model "answering" a question that was actually part of the pasted content.

Worked Examples

Example 1 - Generic Answer → Specific Answer

Before: "Explain machine learning."

After: "Explain machine learning to a marketing manager with no technical background, using a real marketing analogy, in under 150 words."

Example 2 - Inconsistent Format → Reliable Format

Before: "Summarize this article."

After: "Summarize this article in exactly this format: 1) One-sentence thesis, 2) Three key points as bullets, 3) One counterargument the author didn't address."

Example 3 - Shallow Reasoning → Verified Reasoning

Before: "Should we launch this product in Q3 or Q4?"

After: "Think step by step through the tradeoffs of launching this product in Q3 vs Q4, considering [[seasonality, competitor timing, team bandwidth]]. State your recommendation, then list the assumptions behind it so I can check which ones actually hold."

Pattern-to-Use-Case Quick Reference

Common Mistakes to Avoid

• No role specified - "You are a helpful assistant" doesn't narrow behavior at all; a specific role changes vocabulary, tone, and what the model treats as relevant.
• No output format, leading to inconsistent structure across runs, especially painful when you're pasting output into a template or spreadsheet.
• Accepting the first output without one round of specific, targeted iteration - the first response is a starting point, not a final answer.
• Vague feedback ("make it better") instead of specific direction ("cut the intro paragraph, add one data point") - vague feedback gets vague revisions.
• Mixing instructions and content with no delimiter, confusing the model about what's a command vs. what's the input to be processed.

Master these 5 patterns and most "AI isn't giving me good output" problems disappear. Start with whichever pattern maps to your diagnosed symptom above, not with all five at once.