There's a persistent belief that AI tools either work well or they don't — that results are mostly a function of which model you're using. This isn't true. After looking at thousands of prompts, the same five mistakes account for the majority of disappointing outputs. Every one of them is fixable in under two minutes.
Most people open with what they want. "Write a contract review." "Summarize this document." "Create a marketing plan." The problem: without a role, the model defaults to a generalist answering style that's superficially correct and practically useless. OpenAI's prompt engineering guide explicitly recommends persona assignment as the highest-impact single tactic.
The fix is simple — one sentence before your actual request that establishes who the model should be:
The role activates the right domain knowledge and sets the tone and depth of the response.
When you don't specify what you want, the model guesses. And it usually guesses the average answer — the most generic, broadly applicable, inoffensive response. That's rarely what any specific professional needs.
If you actually need a 400-word executive summary in memo format with a risk assessment, say that. If you need three options with trade-offs, ask for three options with trade-offs. Unstated requirements produce outputs that don't meet them.
Before submitting any prompt, ask yourself: "What would make me reject this output as unhelpful?" Then add those requirements explicitly to the prompt.
Almost no professional would take a consultant's first draft and ship it unchanged. But that's exactly what most people do with AI outputs — they read it, feel vaguely dissatisfied, and move on. The model maintains context across a conversation. You can refine, redirect, push back, and build on what it generated — a technique documented in Anthropic's iterative prompt refinement documentation.
Useful follow-up moves after a first response:
- "That's too generic. Give me three specific examples from the [industry] context."
- "Rewrite section two to take a more conservative position on the liability risk."
- "This is 800 words. I need it under 300. Cut everything that isn't essential."
- "Now apply this same analysis to a scenario where [variation]."
There's a difference between task context and situation context. Task context is "review this contract." Situation context is everything that affects what a useful review looks like: Who are you? What kind of company is this for? What's the relationship with the counterparty? What are you worried about specifically? What will you do with this review?
Without situation context, the model can only produce a generic analysis. With it, the output becomes directly applicable to your actual circumstance.
Ask a general legal question and you'll often get a disclaimer-heavy response that ends with "consult a qualified attorney." Ask a medical question and you'll get a reminder that this isn't medical advice. These caveats exist because models are trained to be safe — a design choice documented in OpenAI's InstructGPT paper on RLHF safety training — but they erode the utility of outputs when the professional context is already established.
If you know the caveats apply and don't need to be reminded, tell the model that explicitly. "This is for analytical purposes — I understand this isn't legal/medical/financial advice. Skip the disclaimers and give me the direct analysis."
This one instruction can transform a hedged, unhelpful output into something immediately actionable.
The Pattern Behind All Five Mistakes
Look at all five mistakes and you'll notice something: they're all versions of the same error. They're all about leaving things implicit that should be explicit. AI models don't fill in implicit requirements with what you'd want — they fill them in with the average, the safe, the generic. This is a documented behavior arising from reinforcement learning from human feedback (RLHF), where models are optimized toward broadly acceptable responses rather than contextually ideal ones. Every specification you don't provide is an opportunity for the model to produce something you didn't want.
The prompt engineers who consistently get excellent results aren't doing anything exotic. They've just built the habit of making implicit requirements explicit before they ask.
Need a custom prompt? Try our AI Generator
Describe what you need, pick your AI (ChatGPT, Gemini, or Claude), and watch 3 specialized agents craft, refine, and optimize your prompt. Free, no signup.
Try the AI Generator →Enjoyed this? Get prompts like these weekly.
Top AI prompts every Monday, free. No spam.
For the foundational framework behind what makes prompts work, see Best Practices for Writing Effective AI Prompts. If you're using AI in a professional context, the niche-specific advantages are covered in Why Niche-Specific AI Prompts Outperform Generic Ones.