As AI moves from an experiment to an everyday part of agency work, the quality of the output is becoming a bigger competitive issue. Faster production has little value if the result needs to be rebuilt, fact-checked from scratch or stripped of language that sounds unmistakably machine-made. For agencies, the prompts they use to generate AI content are part of the quality-control process.
Beyond telling a large language model what to produce, the strongest prompts shape how AI approaches a problem, helping to surface uncertainty up front and create useful friction before the first drafts of an idea become client-facing work. Here, members of Forbes Agency Council share the valuable instructions they rely on to ensure AI-generated work is accurate, original and useful.
1. ‘Craft The Best Prompt For Me’
Aside from adding preset personalization that encourages AI to challenge my reasoning and prioritize truth over agreement, I have learned that it’s best not to try and craft perfect prompts myself, especially for complex requests. Instead, I start by giving AI a messy version of what I want to achieve, then ask it to craft the best prompt for me. I have found that iterating on a prompt gets me the output I want much quicker than iterating on a response. – Nikos Lemanis, Luxid
2. ‘Be Brutally Honest’
I’d rather AI tell me my brilliant idea is garbage than the customer. My employees usually won’t, at least not to the degree I need to hear it. If there’s an ounce of good idea deep inside, then help me get it to the surface. If not, tell me to scrap it so I can move on with life. This saves me more time, on a daily basis, than I can calculate. – Michael Chagala, Rank Harvest Digital Marketing
3. ‘Act As An Advisor, Not An Assistant’
In Claude, I have instructed it not to act like my assistant, but as an advisor who is smarter than me and meant to challenge and ask questions until we get a good outcome. The quality difference in results has been incredible. – Mitchell Leiman, iPromote
Forbes Agency Council is an invitation-only community for executives in successful public relations, media strategy, creative and advertising agencies. Do I qualify?
4. ‘Build A Claim-Checking Audit’
The most important thing is to have a fact-checking process. I often use two LLMs (usually Claude and ChatGPT) to do that, or have a “builder” and “checker” when vibe-coding. But the final step always involves asking the LLM to generate a table of every claim in the content, including where to look and where the claim came from, and then flag the order of priority for checking. This forces you to check for hallucinations and inaccuracies. – Matt Wilkinson, Strivenn
5. ‘What Am I Not Seeing?’
One of the prompts I use most is, “What am I not seeing?” Before I rely on AI’s first response, I ask it to identify blind spots, challenge my assumptions and offer perspectives I may have overlooked. It’s a simple prompt, but it consistently leads to stronger thinking and better decisions. I’ve found AI is most valuable when it helps me question my own perspective, not just confirm it. – Jacquelyn LaMar Berney, VI Marketing and Branding
6. ‘Mark Anything You Cannot Verify’
My standing instruction is one line: “Mark anything you cannot verify, and do not fill gaps.” AI fails confidently, which is the dangerous part. Early on, a generated statistic nearly slipped into a client pitch to a journalist, and one invented number can end that relationship for good. Forcing the model to show its uncertainty turns it from a smooth liar into a useful research assistant. – Boris Dzhingarov, ESBO Ltd
7. ‘Don’t Use Any Of The Phrases On This List’
Ever read something that opened with “In today’s rapidly evolving technology landscape…” and instantly knew it was AI? Or you’ve seen “The ones who are winning are doing X, Y and Z”? Those are AI’s fingerprints, and once you spot them, they’re everywhere. Because of this, my go-to instruction is a banned phrase list. Naming the specific patterns that read as generic AI works far better than just asking for “more human writing” after an output. – Parry Headrick, Crackle PR
8. ‘Use Multiple Perspectives To Verify’
AI’s biggest weakness isn’t creating content. It’s sounding confidently wrong. I never rely on a single prompt. Every workflow includes multiple perspectives, iterations, self-audit, source validation and, for critical work, a second AI to challenge the output. Most AI platforms state it can make mistakes. Trust AI. Verify everything. – Monica Alvarez-Mitchell, Pulse Creative, LLC
9. ‘Do Not Invent Missing Details’
I often add: “Separate facts from assumptions, flag anything uncertain and do not invent missing details.” This prevents AI from filling gaps with confident-sounding guesses. It is useful for research and analysis because it makes the output easier to verify. – David Ispiryan, Effeect
10. ‘Ask Me Questions If Context Is Missing’
One instruction I always use is, ‘Ask me any questions before you answer if you don’t have enough context.’ It prevents AI from making assumptions and leads to much stronger results. I’ve learned that the quality of the output depends less on the tool and more on the context you give it. – Solomon Thimothy, Clickx
11. ‘What Would An Expert Disagree With?’
I always end with: “Challenge my assumptions. What’s missing? What would an expert disagree with? Ask me clarifying questions before answering if needed.” AI’s biggest weakness is how it confidently fills gaps. By inviting pushback instead of confirmation, I get more balanced insights, stronger thinking and far fewer generic responses. – Sun Yi, Night Owls
12. ‘Surface The Context Gaps’
Ask the model what it’s assuming and what it would need to know before it answers. Most bad outputs aren’t a result of a failure in reasoning; they’re actually context gaps the model filled in on your behalf. Forcing those gaps into the open turns a confident wrong answer into a short list of questions you can actually go resolve. The common gaps should be part of the everyday context markdowns that you supply the AI with. – Lior Eldan, Moburst
13. ‘What Are We Assuming Here?’
The prompt I use most is simple: “What are we assuming here?” It is not fancy, but it saves bad work. AI is very good at making weak thinking sound finished. In B2B communications, that is dangerous. A wrong assumption about the audience, market, buyer pain or proof point can derail the whole output. This prompt forces the team to challenge the brief before polishing the answer. – Lars Voedisch, PRecious Communications
14. ‘What Details And Data Do You Still Need?’
My go-to fix is feeding the AI deep, out-of-the-box context on our client, competitors and niche first. The secret is reversing the loop at the end by asking: “What specific details or data do you need from me to fill gaps and make this practical?” This command stops generic assumptions, yielding highly original, execution-ready workflows. For regulated fields like healthcare and legal, we restrict AI to verified inputs only to prevent hallucinations and ensure accuracy. – Vin Sonpal, CS Web Solutions
15. ‘List The Three Ways This Answer Could Fail’
My go-to instruction is: “List the three ways this answer could fail before you write it, then correct for them.” It works because most weak output is not wrong in one big way, but in small misses of context, tone or evidence. Surfacing likely failure modes first makes the response more disciplined, less polished-for-show and far more useful in real decision-making. – Vaibhav Kakkar, Digital Web Solutions
16. ‘Show Me What You Are Not Sure About’
This one line changed everything for us. AI defaults to confidence even when it is guessing, and confident guessing is what gets published and embarrasses you in front of a client. Force it to flag its own uncertainty, and the weak spots surface immediately. Half the time, the answer is fine. The other half is exactly where I need to do the thinking myself. – Tessar Napitupulu, Arfadia
17. ‘Avoid Contrast-Based Setups And State Points Directly’
My go-to instruction is: “Avoid contrast-based setups and state the point directly.” I’ve learned that this simple prompt cuts AI filler and makes the output clearer and more natural. It also encourages stronger framing, reduces unnecessary explanation and keeps the response focused on the main idea. – Goran Paun, ArtVersion
18. ‘Ask Me What You Need To Know Before Answering’
I end AI prompts with, “Ask me what you need to know before answering.” It flips the AI from guessing to interviewing, and the questions expose what context I forgot to give—whether it’s the audience, goal or constraints. Most bad AI output isn’t the tool’s fault; it’s a starved prompt. Making it ask first kills the generic slop before it gets generated, instead of me editing it out after. – Nicholas Cormier, Home Builder Marketers
