ChatGPT gives poor results when prompts lack three things: a role, an audience, and constraints. Fix those three and the output quality transforms — without changing the AI, the model, or your subscription. Here's exactly what goes wrong and how to correct it.
You type a question into ChatGPT. It gives you something back. Technically an answer. But it's too vague, too long, weirdly formal, or just not what you needed.
You try again. Same problem.
You start wondering if this AI thing is overhyped.
It isn't. The problem is almost never the AI. The problem is how we talk to it.
ChatGPT Is Not Google
Most people approach ChatGPT the way they approach a search engine — short, keyword-style queries.
"Marketing ideas."
"Email to client."
"Summary of meeting."
Google was trained to interpret fragments. ChatGPT wasn't. It's a reasoning system that works best when you treat it like a highly capable colleague who just started today. Intelligent, willing — but with no context about your business, your audience, your tone, or what "good" looks like to you.
When you give it nothing, it invents context. And it usually invents the wrong one.
The 3 Reasons ChatGPT Gets It Wrong
1. No role
ChatGPT defaults to a generic, helpful assistant voice when you don't tell it who to be. That's why your emails come out formal when you wanted casual, or why your copy sounds like it was written by a committee.
The fix: tell it who it is before you give it the task.
"You are a direct, friendly copywriter who writes for small business owners. Write a follow-up email to a client who hasn't responded in two weeks."
That single change transforms the output.
2. No audience
"Write a LinkedIn post about my new service" produces bland filler because the AI knows nothing about who will read it. Are they senior executives? Freelancers? Sceptics?
Always tell the AI who the reader is. "Write this for a marketing manager who is sceptical of AI tools." Watch how differently it responds.
3. No constraints
Without limits, AI fills space. You ask for a summary and get eight paragraphs. You ask for an email and get a dissertation.
Add a word count, a format, a tone. "Under 100 words, no bullet points, conversational." Constraints aren't restrictions — they're instructions.
A Real Before and After
Weak prompt:
"Give me ideas for social media content."
What you get: a generic list that could apply to any business on earth.
Strong prompt:
"You are a social media strategist. I run a one-person HR consultancy in the UK. My audience is HR directors at companies with 50–200 employees. Give me 5 LinkedIn post ideas that show expertise without sounding like I'm selling. Two sentences each."
What you get: five posts you could publish tomorrow.
The difference is not the AI. It is the prompt.
Why This Matters More Than Which AI You Use
In April 2026, Anthropic published results from an internal experiment called Project Deal. Sixty-nine employees sent Claude agents — acting autonomously as buyers and sellers — into a real marketplace with a $100 budget each. The agents completed 186 deals worth over $4,000 in real goods, with no humans in the loop during negotiations.
The finding that stood out: the quality of the instructions given to the agent determined the economic outcome. Better-briefed agents consistently outperformed worse-briefed ones — not by a small margin, but measurably, deal after deal.
This isn't a ChatGPT story. It's a prompting story. And it scales from a negotiation experiment all the way down to the email you're asking AI to write this afternoon.
Where to Go From Here
Role, audience, constraints — these three fixes are the foundation. There are techniques that go further: giving AI examples of your own writing to match your voice, chaining tasks instead of combining them, asking it to argue back at you to catch mistakes early.
None of it requires technical knowledge.
Unit 1 of Agent Tongue is completely free — no account required, no credit card. It covers the foundations of prompting in around an hour and makes you demonstrably better by the end, not just better informed.
If it's not for you, you've lost nothing. If it is, Units 2–8 cover everything from advanced techniques to AI agents to prompt security — one-time £39, no subscription.
Sources: Anthropic Project Deal research paper, published April 2026.
Prompt examples above are illustrative of technique, not attributed to specific organisations.