AI for Work: Designing good prompts for real work

Designing Good Prompts for Real Work

What This Is

Prompt design is the skill of crafting clear, structured inputs to guide AI models (like LLMs) to produce useful, accurate outputs for work tasks. Poor prompts waste time and money; well-designed ones save hours, reduce errors, and unlock AI’s full potential. Example: A marketing manager uses a poorly worded prompt ("Write a blog post about AI") and gets a generic, off-brand draft. With a refined prompt ("Write a 500-word blog post for tech-savvy small business owners explaining how AI can automate customer support, using a friendly but professional tone and including two real-world examples"), they get a targeted, publish-ready draft in one try.

Key Facts & Principles

• Clarity over creativity: State the task, audience, and constraints upfront. Example: Instead of "Help me with my email," use "Draft a polite but firm email to a vendor who missed a deadline, emphasizing our contract terms and requesting a revised timeline by EOD Friday."
• Role prompting: Assign the AI a specific role to shape its output. Example: "Act as a senior data analyst. Explain the limitations of this regression model to a non-technical executive in 3 bullet points."
• Context window: Most models have a limited "memory" (e.g., 4K–32K tokens). Prioritize recent/relevant info. Example: For a long document, summarize key points first, then ask the AI to analyze only those.
• Temperature: Controls randomness (0 = precise, 1 = creative). Use 0.2–0.5 for factual tasks (e.g., reports), 0.7–0.9 for brainstorming. Example: "Generate 5 tagline ideas for our new product (temperature: 0.8)."
• Chain-of-thought (CoT): Ask the AI to "show its work" to improve accuracy. Example: "Explain step-by-step how you’d calculate customer churn rate for Q2, then provide the final number."
• Few-shot learning: Provide 1–3 examples of desired output to guide the model. Example: "Here are two strong subject lines for our newsletter: [Example 1], [Example 2]. Write 3 more in the same style."
• Negative prompting: Explicitly exclude unwanted elements. Example: "Summarize this meeting transcript in 3 bullet points, excluding off-topic discussions about office snacks."
• Iterative refinement: Treat prompts like code—test, debug, and optimize. Example: Start with a broad prompt ("Analyze this dataset"), then narrow based on the output ("Focus on outliers in the ‘revenue’ column and suggest causes").
• Bias mitigation: Avoid leading questions. Example: Instead of "Why is our competitor’s product better?" ask "Compare the strengths and weaknesses of our product vs. Competitor X’s."
• Output format: Specify structure (e.g., table, bullet points, JSON) to save post-processing time. Example: "List the top 5 risks in this project plan as a markdown table with columns: Risk, Likelihood (1–5), Mitigation."

Step-by-Step Application

1. Define the goal
2. Ask: What’s the exact output I need? (e.g., a decision, a draft, a summary).

3. Example: "I need a 1-page executive summary of this 20-page report, highlighting key findings and recommendations for the board."

4. Gather context

5. Include:
   • Audience (e.g., "for a non-technical client").
   • Tone (e.g., "professional but approachable").
   • Constraints (e.g., "under 200 words," "avoid jargon").

6. Example: "Write a LinkedIn post announcing our new feature. Audience: SaaS founders. Tone: excited but not salesy. Length: 150–200 words. Include a CTA to book a demo."

7. Structure the prompt

8. Use this template: Role: [Who should the AI be?] Task: [What should it do?] Context: [Relevant background] Constraints: [Word count, format, exclusions] Examples: [1–2 samples of desired output, if helpful]

9. Example: Role: Act as a product manager with 10 years of experience in fintech. Task: Draft a product requirements document (PRD) for a new fraud-detection feature. Context: Our users are small-business owners who lose $5K/month on average to fraud. Current tools are too complex. Constraints: 1-page max, use bullet points, avoid technical terms like "ML models." Examples: [Attach a PRD snippet from a past project.]

10. Test and refine

11. Run the prompt, then ask:
    • Is the output usable as-is, or does it need tweaks?
    • What’s missing or incorrect?

12. Example: If the AI’s draft is too vague, add: "Include specific metrics (e.g., ‘reduce false positives by 30%’) and a timeline for implementation."

13. Add guardrails

14. For high-stakes tasks, include:
    • Verification steps: "Cross-check these calculations with the attached spreadsheet."
    • Fallbacks: "If you’re unsure about a fact, say ‘I don’t have enough data’ instead of guessing."

15. Example: "Summarize this legal contract, but flag any clauses that seem unusual or risky. Do not interpret the law—just highlight areas for our lawyer to review."

16. Automate (if repeated)

17. Save prompts as templates (e.g., in a shared doc or tool like Notion).
18. Example: Create a "Prompt Library" with templates for:
    • Meeting summaries
    • Email responses
    • Data analysis requests

Common Mistakes

• Mistake: Overloading the prompt with irrelevant details. Correction: Prioritize the last 3–5 sentences of context—models pay most attention to recent input. Why: Extra details dilute focus and waste tokens.

• Mistake: Assuming the AI "knows" your company’s style or data. Correction: Attach or describe your brand guidelines, past examples, or datasets. Why: Models default to generic outputs without context.

• Mistake: Using vague verbs like "analyze" or "improve." Correction: Replace with specific actions: "List the top 3 trends in this dataset with supporting data points." Why: Vague verbs lead to vague outputs.

• Mistake: Ignoring the output format until after generation. Correction: Specify format in the prompt (e.g., "as a table," "in JSON," "as bullet points"). Why: Reformatting manually wastes time.

• Mistake: Treating the first output as final. Correction: Plan for 2–3 iterations. Why: Even great prompts often need tweaks for edge cases.

Practical Tips

• Use "prompt chaining" for complex tasks: Break work into steps. Example:
• "Summarize this customer feedback into 5 key themes."
• "Prioritize these themes by frequency and impact."

• "Draft a response to the top theme, addressing the core concern."

• Leverage "prompt engineering" tools: Use tools like PromptPerfect or SnackPrompt to refine prompts before running them.

• Document your prompts: Keep a log of what worked (and what didn’t) for future reference. Example: "Prompt for Q2 financial summary: 80% accuracy, needed 2 iterations to fix revenue numbers."

• Watch token costs: Long prompts + long outputs = higher costs. Trim fluff and use concise language. Example: "Summarize this 10-page doc in 200 words" vs. "Read this doc and tell me what it says."

Quick Practice Scenario

Scenario: You’re a project manager. Your team just finished a sprint, and you need to update stakeholders. The sprint notes are messy, with 20+ bullet points, off-topic discussions, and no clear takeaways. Write a prompt to generate a concise update email.

Answer: "Act as a senior project manager. Review these sprint notes [paste notes] and draft a 150-word email to stakeholders summarizing: - 3 key accomplishments (with 1-sentence impact statements). - 2 blockers (and next steps to resolve them). - The goal for the next sprint. Use a professional but warm tone. Exclude off-topic discussions (e.g., team lunch plans)."

Explanation: The prompt assigns a role, specifies output format, excludes noise, and sets a word limit—all critical for a usable draft.

Last-Minute Cram Sheet

1. Clarity > creativity: State the task, audience, and constraints upfront.
2. Role prompting: "Act as a [role]" shapes output tone and depth.
3. Token limit: 1 token-0.75 English words; prioritize recent context.
4. Temperature: 0.2–0.5 for facts, 0.7–0.9 for ideas.
5. Chain-of-thought: "Explain your reasoning step-by-step" improves accuracy.
6. Few-shot learning: Provide 1–3 examples of desired output.
7. Negative prompting: "Exclude [X]" avoids unwanted content.
8. Output format: Specify structure (table, JSON, bullets) in the prompt.
9. Avoid leading questions: "Why is X better?"-"Compare X and Y."
10. Iterate or fail: Plan for 2–3 prompt refinements per task.