The mental model that fixes most prompt problems
Most "bad" AI output isn't actually the AI's fault. It's a prompt problem.
The mental shift that fixes 80% of prompt issues: stop thinking of prompts as questions you're asking the AI. Start thinking of them as job descriptions you're writing for a competent contractor.
When you hire a contractor, you don't say "write me a marketing email." You say:
"Write a marketing email to our existing customers (mostly busy restaurant owners) announcing our new Lightning checkout integration. Keep it under 200 words. Lead with the time savings, not the technology. Don't use the word 'revolutionary.' End with a one-line CTA to book a 30-minute demo. Match the voice of [examples attached]. Avoid these phrases [list]."
That's a job description. The AI handles it well because there's actually enough specification to handle it well. "Write me a marketing email" fails because the contractor — human or AI — has to guess at every dimension you didn't specify.
The skill of prompt writing is really the skill of specifying what you want with enough detail that a competent worker could deliver it without coming back for clarification.
The structure that works
A prompt that produces reliable output covers five elements:
1. Role. Who is the AI being asked to be? "You're a senior copywriter who specializes in B2B SaaS" or "You're a customer service representative for a dental practice." Setting the role primes the AI to use the vocabulary, tone, and reasoning patterns of that role.
2. Task. What are they doing? The verb matters — "write," "summarize," "categorize," "extract" all produce different output shapes.
3. Context. What does the AI need to know that isn't already obvious? Audience, business model, customer details, prior interactions, constraints. The information a new hire would need to do this job competently.
4. Format. How should the output be structured? "Three bullet points" or "A 200-word email with subject line" or "JSON with the fields: name, category, amount, date."
5. Constraints. What should the AI not do? "Don't use the word 'leverage'" or "Keep it under 150 words" or "Avoid em-dashes." Negative constraints are often more important than positive instructions.
Bad prompt vs. good prompt
Bad prompt: "Write me an email to a customer who hasn't paid their invoice in 60 days."
What you get: a generic, possibly aggressive, possibly groveling email that doesn't match your brand voice, your customer relationship, or your actual situation.
Good prompt:
Role: You're a customer success manager at a B2B software company.
Task: Write an email to a customer who has an outstanding invoice 60 days overdue.
Context: The customer is a long-time client (3+ years), the invoice is $4,500 for our monthly service, and we've already sent one polite reminder at 30 days. We value the relationship but need to resolve this.
Format: A professional email under 150 words, with a clear subject line. No bullet points.
Constraints: Don't be threatening. Don't be apologetic. Don't mention legal consequences. Use 'we' language, not 'I.' End with an offer to discuss the situation, not just a payment demand.
The difference isn't that the second prompt is more clever. It's that the second prompt has done the thinking that the first prompt left to the AI to guess at.
The three iteration techniques
Technique 1: "More like X, less like Y." Name reference points the AI knows — companies, writers, brands. "More like Notion's voice (clean, confident, direct) and less like Mailchimp's (chirpy, exclamation-heavy)."
Technique 2: "Show me three different versions." When you're not sure what you want, ask for variety. Three options often produces better thinking than asking for "the right version."
Technique 3: "Critique your own draft." "Critique this draft as if you were a senior editor reviewing junior work. What are its weaknesses? Then rewrite it addressing them." Forcing self-critique often beats asking "make it better."
The error patterns to recognize
AI is too generic. Add more specific context about your business, audience, voice.
AI is too verbose. Specify length explicitly ("under 100 words," "exactly 3 bullet points").
AI invented facts. Include factual context in the prompt and add "Only use facts I've provided; don't invent details."
AI used wrong tone. Name a tone explicitly with examples.
AI ignored an instruction. State constraints at the end of the prompt — AI gives more weight to the last instructions it reads.
Each of these is fixable in the prompt. None of them are fixed by switching to a different AI model.
When prompts aren't enough
Some tasks can't be handled by a single prompt — tasks needing persistent context across sessions, specific tool access (CRM, database), actions in the real world (actually sending the email), or multi-step reasoning with verification.
These are the cases where you've outgrown simple prompting and need either an agent (Lesson 17) or a custom deployment (Lesson 18). For most operator work, though, well-structured prompts handle 90% of what you need.
What's next
You now know the mental model (prompts as job descriptions), the five-element structure, three iteration techniques, the error patterns that signal prompt problems vs. AI problems, and when you've outgrown prompting.
Next up: Lesson 20 — When to put a human in the loop.
Frequently asked
Questions that come up after this lesson.