Prompting has become a go-to tool for teams experimenting with AI, but many soon discover a recurring issue: inconsistency. Ask the same question twice and you might get different answers. Change the phrasing slightly, and the tone or structure can shift. For individuals, that’s frustrating. For businesses, it undermines confidence and limits uptake.
When teams can’t rely on AI to deliver consistent, high-quality results, it creates a barrier to using it. When you add in the possibility of hallucination (where AI produces entirely false results), then you can understand why a business might hesitate to do anything meaningful with this technology.
So how do we fix it?
Step One: Better Prompts
Before you introduce new layers of complexity, you need to ensure your base prompts are solid. The best prompts share five common traits:
- Clear context – What’s the task and why does it matter?
- Defined role – Who should the AI act as? (e.g. “Act as a financial analyst.”)
- Specific instructions – What exactly do you want done?
- Constraints – Format, tone, word count, key points to include.
- Expected outcome – What does success look like?
Here’s a before and after example:
Before:
“Summarise this customer research.”
After:
“Act as a senior customer insights analyst. Summarise the attached customer research for the CMO of a SaaS company. Focus on the top 3 actionable insights. Include supporting data. Use a concise format with clear section headings. Keep it under 400 words and prepare it in a style suitable for pasting into an executive slide deck.”
The difference is night and day. The “latter”after” has a sharper focus, better structure, and far more usable output.
Please note that even the “after” example above is fairly simplistic. I’ve seen several really “next level” examples which are multiple pages long, split tasks into multiple phases, ask the AI to show reasoning, and provide complex examples and edge cases.
Step Two: Meta Prompting
This sounds complicated and even more “nerdy”, but it’s actually quite simple.
The best “person” to write the most effective prompt, is the one who will be executing it – i.e. the AI.
I must confess this is something I never thought of until recently. When I first heard of “meta prompting”, and I realised the potential, my mind was blown 🤯
Using this approach we start with our initial prompt, run through several iterations, provide feedback, and then ask the AI to generate the most effective prompt to achieve the desired output.
Step Three: Refine and Expand
From here, you can keep going. Add layers. Introduce review cycles. Ask the AI to explain its reasoning.
“After generating your output, explain how you structured it. Identify any areas where you lacked information or had to make assumptions. If unsure, state clearly where more input is needed.”
This ‘trap door’ – where the AI can admit it doesn’t have enough information is key. It reduces the risk of hallucinations and lets you know when human input is required.
Prompts can be long. The best ones often are. That’s fine. These are repeatable assets, not one-off hacks. Once you’ve created a strong prompt for a particular task whether it’s summarising customer feedback, preparing a competitor snapshot, or building a draft proposal you can re-use, tweak, and systematise it across teams.
Wrapping Up
If you can see the potential of implementing prompting in your company, but are concerned about a lack of control, the potential for variances in output, or just plain hallucinations – then following the steps above should provide you with a far more solid set of prompts. Test continually, tweak, and as you achieve consistency, you’ll also gain confidence.
