AI Literacy: Prompts instructions and constraints

Prompts, Instructions, and Constraints: A Practical Study Guide

What This Is

Prompts, instructions, and constraints are the input controls you use to guide AI models toward useful, accurate, and safe outputs. They matter in everyday work because poorly designed inputs lead to vague, off-brand, or risky results—while well-structured ones save time, reduce errors, and align AI with business goals. Example: A marketing team uses a prompt like "Write a 150-word LinkedIn post in a professional but conversational tone about our new SaaS tool for HR teams. Avoid jargon. Include a CTA to book a demo." instead of "Write something about our product." The first version yields publishable content; the second requires heavy editing.

Key Facts & Principles

• Prompt: The text input you give an AI to elicit a response. Example: "Summarize this 10-K filing in 3 bullet points for a non-finance audience."
• Instruction: A clear directive within a prompt that tells the AI what to do. Example: "Translate this email into Spanish, but keep the tone formal."
• Constraint: A rule that limits or shapes the AI’s output (e.g., length, format, tone, exclusions). Example: "Do not mention competitors by name. Use no more than 200 words."
• Zero-shot vs. few-shot prompting:
• Zero-shot: No examples provided (e.g., "Explain quantum computing like I’m 5.").
• Few-shot: Includes 1–3 examples to guide output (e.g., "Here are two product descriptions I like. Write a third in the same style.").
• Role prompting: Assigning the AI a persona to improve relevance. Example: "Act as a cybersecurity consultant. Explain phishing risks to a board of directors."
• Chain-of-thought (CoT): Asking the AI to "show its work" to improve reasoning. Example: "Before answering, list the steps you’d take to analyze this dataset for anomalies."
• Temperature: A setting (0–1) controlling randomness. Low (0.2) = predictable; high (0.8) = creative. Example: Use 0.2 for legal summaries, 0.7 for brainstorming taglines.
• Hallucination mitigation: Constraints that reduce false outputs. Example: "Only use data from the attached PDF. If unsure, say ‘I don’t know.’"
• Output formatting: Specifying structure to save post-processing time. Example: "Return the analysis as a Markdown table with columns: Risk, Impact, Mitigation."

Step-by-Step Application

1. Define the goal
2. Ask: What problem am I solving? (e.g., draft an email, analyze data, generate ideas).

3. Example: "I need a concise summary of customer feedback to present to leadership."

4. Write the base prompt

5. Start with a clear instruction + context.

6. Example: "Summarize the key themes from these 50 customer survey responses. Focus on pain points and feature requests."

7. Add constraints

8. Limit scope, tone, length, or exclusions.

9. Example: "Use bullet points. Keep it under 200 words. Avoid generic phrases like ‘great product.’"

10. Include examples (if needed)

11. For complex tasks, provide 1–2 examples of desired output.

12. Example: "Here’s how I’d summarize feedback: ‘- 60% want faster load times (current avg: 4.2s). - 25% ask for dark mode.’"

13. Set parameters

14. Adjust temperature, max tokens, or response format.

15. Example: "Temperature: 0.3. Output as a numbered list."

16. Test and iterate

17. Run the prompt, evaluate output, and refine.
18. Example: If the summary is too vague, add: "Prioritize the top 3 themes with supporting data."

Common Mistakes

• Mistake: Overloading a single prompt with multiple tasks.

• Correction: Split into smaller prompts (e.g., "First, extract quotes. Then, summarize themes."). Why? Models perform better with focused tasks.

• Mistake: Assuming the AI "understands" implicit context.

• Correction: Explicitly state constraints (e.g., "Do not include internal project names."). Why? AI lacks real-world knowledge unless told.

• Mistake: Ignoring output formatting until after generation.

• Correction: Specify format upfront (e.g., "Return as a JSON object with keys: ‘theme’, ‘count’, ‘example’."). Why? Saves reformatting time.

• Mistake: Using vague language (e.g., "Make it better").

• Correction: Define criteria (e.g., "Rewrite this email to be 30% shorter and more action-oriented."). Why? Quantifiable instructions improve results.

• Mistake: Skipping few-shot examples for complex tasks.

• Correction: Provide 1–2 examples for nuanced work (e.g., legal drafting, technical writing). Why? Reduces trial-and-error.

Practical Tips

• Start with a template: Save time by reusing prompt structures (e.g., "Summarize [document] for [audience] in [format] with [constraints].").
• Use "negative prompts": Explicitly exclude unwanted outputs (e.g., "Do not use metaphors or idioms.").
• Leverage role-playing for tone: Assign a role to match brand voice (e.g., "Write as a friendly but authoritative customer support agent.").
• Automate constraints: For repetitive tasks, use tools like Zapier or custom scripts to pre-fill prompts (e.g., always add "Max 150 words" to social media prompts).

Quick Practice Scenario

Scenario: You’re a product manager drafting release notes for a software update. The AI’s first draft is too technical for end users. Question: How would you revise the prompt to get a more user-friendly version?

Answer: "Rewrite these release notes for non-technical users. Explain each feature in 1–2 sentences using plain language. Avoid jargon like ‘API,’ ‘latency,’ or ‘backend.’ Focus on benefits (e.g., ‘faster load times’ instead of ‘optimized database queries’)." Explanation: Constraints on tone, language, and focus shift the output from developer-centric to user-centric.

Last-Minute Cram Sheet

1. Prompt = instruction + context + constraints.
2. Constraints reduce hallucinations and off-brand outputs. Don’t assume the AI "knows" your limits.
3. Few-shot > zero-shot for complex tasks. Examples improve accuracy.
4. Low temperature (0.2–0.4) for facts; high (0.7–0.9) for creativity.
5. Role prompting improves relevance (e.g., "Act as a [job title].").
6. Chain-of-thought (CoT) boosts reasoning (e.g., "Explain your steps first.").
7. Output formatting saves post-processing time (e.g., "Return as a table.").
8. Negative prompts exclude unwanted content (e.g., "No emojis.").
9. Split multi-step tasks into separate prompts. Avoid "do everything" prompts.
10. Test prompts with small samples before scaling. Iterate to avoid costly errors.