Most business prompts read like telegrams to a genie. “Write a sales email.” “Summarise this.” “Make it better.” Then we’re surprised when the output is generic, misses the point, or needs three rounds of fixing before it’s usable.
The problem isn’t the model. It’s the instruction. And the fix is something every business owner already knows how to do: write a job description.
A Prompt Is a Job Description, Not a Wish
When you hire someone, you don’t say “do the marketing” and walk away. You write a job description. You specify the role, the context, the constraints, the deliverable, and what good looks like. You give examples when the task is subtle. You name the things to avoid.
A prompt is the same document, compressed into a message. The people getting the most out of AI in business aren’t the ones with the cleverest prompts or the longest ones. They’re the ones who instinctively transfer the discipline of writing a good job description into the chat box.
What a Job Description Has That a Wish Doesn’t
Think about the last role you hired for. The job description probably had six things in it. Your prompts should have most of them too.
- Role. Who is this person? “You are a B2B SaaS copywriter who writes for technical buyers.” Not “write copy.” The role sets the voice, the audience, and the assumptions the AI will make about what counts as good.
- Context. What’s the situation? “We’re a 12-person analytics company. The prospect is a COO at a mid-sized logistics firm who asked about pricing after a demo.” Context is what stops the output from sounding like it was written for nobody.
- Task. What exactly should they produce? “Write a 120-word follow-up email that acknowledges her pricing question, offers a specific next step, and doesn’t oversell.” One task, one deliverable.
- Constraints. What are the rules? “No exclamation marks. No ‘revolutionary’ or ‘game-changing.’ Short sentences. One call to action.” Constraints are where most of the quality comes from.
- Examples. Show one email you’ve sent before that landed well. Paste the tone you want. Examples do more than any adjective.
- What to avoid. The things the AI reliably gets wrong if you don’t name them. “Don’t invent features. Don’t mention competitors by name. Don’t use the word ‘synergy.’”
You don’t need all six every time. A quick internal note might only need role, task, and constraints. But every time you’re unhappy with an AI output, the first question isn’t “is the model broken?” It’s “which of these six did I leave out?”
Before and After
Here’s the difference in practice.
Wish: “Write a follow-up email to a prospect who went quiet.”
Job description: “You’re a B2B sales assistant. A COO at a logistics company saw our demo two weeks ago, asked about pricing, and hasn’t replied since. Write a 100-word follow-up email that references the pricing question specifically, suggests a 15-minute call to compare options, and sounds like a person, not a template. No urgency language. No ‘just checking in.’ One clear call to action.”
The first prompt gives you a generic nudge. The second gives you something you can send with minor edits. Same model. Different instruction.
The Three Mistakes That Waste the Most Time
1. Asking for too much at once
“Write a blog post, then create a social caption, then draft three email variants, then suggest a headline.” That’s four jobs. A job description describes one role with one deliverable. When you stack tasks, the AI averages across them and every output gets worse. Split it into separate prompts, take the output of one as input to the next, and the quality jumps.
2. Describing the output instead of the job
“Make it professional but friendly, concise but detailed, persuasive but not pushy.” This is you describing the texture of the output you want, which the model can’t reliably hit because those adjectives mean different things in different contexts. Instead, describe the job: who is writing this, to whom, for what purpose. The tone follows from the role. A customer support agent writes differently from a sales rep. You don’t need to specify the tone if you specify the person.
3. Never writing down what you fixed
Every time you edit an AI draft before sending it, you’re telling the model something it didn’t know. But if that fix lives only in your head, you’ll make the same edit next week. The teams whose AI output keeps improving are the ones who capture the recurring corrections and fold them back into the prompt. That’s not a prompt trick. That’s a feedback loop. And it’s the difference between an assistant that runs in place and one that compounds.
Why This Compounds
Prompt craft isn’t a one-time setup. It’s an operating discipline, the same way hiring is. You don’t write a job description once and never revisit it. You refine it after the first month, after the first big mistake, after the first time someone does the job better than you expected. Prompts deserve the same treatment.
The practical version is simple. Keep your prompts in a shared place, not buried in chat history. When someone on your team writes a prompt that produces great output, save it. When a prompt produces something that needed heavy editing, add a line about what was missing. Over a few weeks, you build a small library of job descriptions that your whole team can reuse. That library is worth more than any single model upgrade.
This is also why we built meLink prompts the way we did. The point isn’t to collect clever prompt templates. It’s to make the job-description discipline reusable — so a good instruction becomes a shared asset your team builds on, not a private trick that leaves when someone goes on holiday.
The One Test
Before you hit send on your next prompt, read it back and ask yourself one question: could a competent stranger do a good job with only this?
If the answer is no, the model probably can’t either. Add the role. Add the context. Name the constraints. Show an example. Then send it.
You’ve got this.
Leave a Reply