AI is wrong sometimes, sure. What worries me more is how often it agrees with you. Ask an out-of-the-box chatbot whether your new pricing is a good idea and it will find nine reasons you're brilliant. That's a mirror, and a mirror can't advise you. You can't run a business on flattery.
So I build honesty into my AI's standing instructions, rules that apply to every conversation before I've typed a word. These are the ones that matter most, straight from the setup I taught live this week.
“AI has a tendency to agree with you.”
The honesty rules
- Always give pros AND cons. I never want a one-sided answer. I want the rounded one, so I'm deciding with open eyes.
- Be opinionated, and be brutally honest when I ask for advice. An AI that hedges everything is useless in a business that requires decisions.
- Play devil's advocate at the end of everything. People say AI keeps you from thinking. This rule does the opposite. It forces me to weigh a real counterargument and make the call myself.
- Channel the top experts on the topic, real ones, including where they disagree with each other. Disagreement between experts is information.
How it works
Two sets of standing rules
So it stops flattering you
Honesty rules
Pros and cons
Be brutally honest
Devil's advocate
Channel the top experts
So it stops making things up
Anti-hallucination rules
Fact-check with sources
Say comedy alert
list what was invented
Open with a code word
A colleague, not a cheerleader
How I do it
Claire's honesty instruction stack
- 1
Pros and cons, always
Never a one-sided answer
- 2
Be brutally honest
Opinionated when asked for advice
- 3
Play devil's advocate
At the end of everything
- 4
Fact-check and cite
Sources with links, checked like an editor would
- 5
Comedy alert
Confess anything invented or unverifiable
Standing rules that apply before she types a word, so the AI argues back instead of flattering her.
Source: Claire B. Soares, from her own travel business and live trainings.
The anti-hallucination rules
- Fact-check every answer with an agent, the way a newspaper fact-checker would, and cite sources with links.
- If anything is made up or can't be verified, say "comedy alert" and list exactly what was invented. Making the AI confess is far cheaper than discovering it later in front of a client.
- Start every chat with a code word. If my AI ever opens without it, I know it's stopped following my instructions. It's my canary in the coal mine.
Why this matters
With these rules in place, my AI stops cheering me on and starts acting like a colleague. It argues back and shows its sources, and it admits what it doesn't know. That's an AI you can actually bring into client work. My complete instruction set, word for word, is inside the free Travel AI Kit. Copy it, then make it yours.
A note before you try this
Test your rules before you trust them. Ask your AI something you already know the answer to, and see if it argues back or calls a comedy alert. Watch for the code word at the start of every chat, because if it's missing, your rules aren't being followed.
Set it up with me
I walk through my standing instructions live in the free AI Masterclass. Bring your own setup and we'll make it honest together.
Common Questions
Why does AI always agree with me?
Chat models are tuned to be agreeable, which makes them poor business advisors by default. The fix is standing instructions that require pros and cons, a devil's advocate position, and cited sources on every answer.
What is a hallucination flag?
It's an instruction that makes the AI announce when it has invented or can't verify something. Claire's version asks it to say 'comedy alert' and list exactly what was made up.
How do I know my AI is still following my instructions?
Ask it to open every chat with a code word. If the code word stops appearing, your instructions are no longer being applied.
Written By
Claire B. Soares
Claire B. Soares is the founder of Travel AI University, a 6x Condé Nast Traveler Top Travel Specialist, and a Certified AI Consultant with a B.S. in Computer Science. A top 10% enterprise sales executive, she sold $3.4B+ in enterprise software to Google, Microsoft, IBM, Boeing, the federal government, and the DoD, then built $16M+ in leisure travel sales.
About Claire
