From the desk of Claire B. Soares
Your AI Is a Yes Girl
And that is only the first reason it feels inconsistent and inaccurate. Where to actually start, with four hours a week and nobody technical to call.
By Claire B. Soares

Every time I tell someone I am an AI expert, I get the same question in about a dozen different ways.
They usually start by telling me it is inconsistent. And that it is not accurate. That the same question gets a brilliant answer on Tuesday and a wrong one on Thursday, and they cannot tell you why.
And then, a little quieter, comes the real question.
“Where do I actually start? And how do I make it work for me?”
So I sat down and wrote the answer. This is it. No cost, no form, no funnel.
But before the answer, you need to meet the reason.
Meet your yes girl
Here is what nobody says out loud.
Most AI will agree with you about ninety nine percent of the time.
Ask it if your idea is good and it will tell you your idea is wonderful. Ask it to review your pricing and it will find something lovely to say about your pricing. Ask it whether you should launch in September and it will help you launch in September, and it would have helped you launch in March with exactly the same enthusiasm.
You walk away feeling great and knowing nothing.
That is not a smart assistant. That is a yes girl. And a yes girl is a very pleasant way to make a bad decision quickly.
I hate yes people. You need a real consultant, not someone who agrees to agree. So I make mine argue with me.
In my settings, I tell it to play devil's advocate on everything. Every answer, every time, whether I asked for it or not.
The effect is immediate. It stops flattering me and starts working for me. And more importantly, it forces me to make a decision, which is the actual point of asking in the first place.
If you only take one thing from this entire piece, take that one. Go add it right now. I will wait. 🥂
The yes girl is a symptom. Here is the disease.
Now, why does your AI behave that way in the first place?
Because you never told it not to.
When people tell me AI is unreliable, they are almost always skipping the same thing. The setup and the personalization.
Go into your settings in Claude or ChatGPT. There is a personalization section sitting right there and most people have never opened it. That is where you set the standard operating procedure for how the thing works with you. Not for one chat. For every chat, forever.
Mine tells it to be opinionated. To give me pros and cons. To format answers professionally. I tell it I have ADHD, so keep it brief and to the point. I tell it to ask me clarifying questions so I can answer in voice mode and we get a better result. I tell it to channel the top experts on whatever I am asking about, and to include experts who disagree with each other.
Then I tell it two more things that do the heavy lifting.
Fact check me like a New York Times fact checker. If it is not verified, it does not get presented as true.
Quote your references, as one of my BFFs Julia says. Give me the links so I can go read it myself.
Here is the thing about accuracy that took me too long to learn. It is not a feature you get. It is an instruction you give. Nobody tells it to verify anything, and then everybody is shocked when it makes something up with total confidence.
If you have not set any of this up, you can genuinely get lost in the sauce. And then it feels like the tool is broken, when really you never told it how you work.
Your AI is not inconsistent and it is not inaccurate. Your setup is. Same question, different day, different context, different answer.
I have packaged the exact instructions I use, word for word, in my AI Starter Kit. It is free. Take it and paste it in.
The butter knife
There is a second reason your AI is inconsistent and inaccurate, and this one has nothing to do with settings.
Think about answering your email.
Most tools will draft a message for you now. Your inbox will offer it. And the draft will be fine, and it will also be wrong, because it does not know your policies. It does not know your products. It does not know your terms and conditions or the way you actually talk to a client who is nervous about a payment.
That is using a butter knife to cut a steak. Can you do it? Yes. Is it hard and slow and a bit undignified? Also yes.
Now imagine you gave it your trips, your services, your terms and conditions, and access to your real sent emails. Now it knows the difference between a client asking about a room category, a client asking about a balance due, and a vendor asking you for something.
Answers that sound like you, flow like you, and are accurate to your business policies with little revision. That is a steak knife.
The gap between those two experiences is not the software. It is about forty minutes of setup that nobody wants to do.
So let's do the forty minutes. Four hours a week, four weeks. Here is the order.
Week one: fix the settings, then open the connectors
Start by downloading Claude Cowork so it lives on your computer, not just in a browser tab. That is my link, and the first three people get a week free.
Then open your settings and write your personalization instructions. Not a prompt. Instructions. Who you are, how you like to be spoken to, what you want it to do before it answers you.
Steal mine if you like. Opinionated. Pros and cons. Brief and to the point. Ask me questions before you answer. Fact check like a New York Times fact checker. Quote your sources. Play devil's advocate at the end of everything.
And add this one, because it has saved me more times than I can count. I tell mine that if it makes anything up or cannot verify something, it has to say comedy alert and then list exactly what it invented.
So when it hands me an answer with a made up statistic in it, it tells on itself. At the bottom. Every time. That single line has caught more errors than any amount of me squinting at a paragraph wondering if it is true.
Then do the part everybody forgets
Go into connectors and plug it into the places your business actually lives. This is the difference between an AI that guesses about you and one that knows.
Here is what mine is connected to and why.
- Gmail
- So it can see what a client actually confirmed, instead of me digging for it.
- Google Calendar
- So it knows what my week looks like before it suggests anything.
- Google Drive
- So it reads my real folders, not a copy I pasted into a chat at eleven at night.
- Fireflies
- So it has my call transcripts. This one is quietly the most powerful. Everything a client tells me on a call is now searchable.
- Notion
- My employee handbook, our processes, the way we do things.
- Stripe
- So the money questions have a real answer behind them.
- Slack
- So team conversations are context, not a separate universe.
- Canva and Airtable
- Design and data, where they already sit.
You do not need all of those. You need the two or three where your business actually happens. For most people that is email, calendar and wherever your files live. Start there.
That is your first hour or two. You will never do it again and it will change every conversation after it.
Week two: pick one job and give it your folder
One job. Not four. The one you resent.
Not the most important one. The one you would pay a stranger fifty dollars to never do again. The follow up you never send. The proposal you rebuild from scratch every time. The captions. The notes after every call.
Then, before you ask it for anything, make it a folder.
If you are working on marketing, that folder should have sample Instagram posts. Sample blogs. Sample emails. A document about your brand voice and how you sound.
Do not have those? Ask it to build them. Point it at your website, point it at your sent folder, and have it write your brand voice document for you. Then read it and fix what it got wrong.
Starting from a folder full of your real work versus starting from a blank chat is night and day. And if you start from the blank chat, I will tell you now what you are going to get. Something that is not quite what you needed, with a few details in it that were quietly invented.
Week three: let it interview you, then write the steps down
This is the part I would put on a billboard.
Go into voice mode and ask it to interview you. Tell it to ask you at least ten questions about what you are trying to do and why, before it produces anything.
Most people type a sentence and hope. Talk to it instead, the way you would talk to a consultant you just hired. You would never say to a real consultant, “make me a marketing plan,” and then go quiet. You would tell them everything. The launch, the four destinations, the hundred thousand dollars you are trying to sell, the client who always books late.
So tell the AI all of that. Out loud, because you will say more out loud than you will ever type. Then tell it to keep asking until it is ninety percent sure it can give you what you want.
This is the heart of my V.O.I.C.E. framework, and it is the single change that moves people from disappointed to dangerous.
Then write the steps down.
This is the honest answer to how long it took me. I have been in this since 2023, working alongside people who do it for a living, and it still did not click into something I could rely on until I slowed all the way down and worked on one process at a time, writing out the steps as I went.
Not a better prompt. Not a better tool. One process, written down, so Thursday gets the same treatment Tuesday got.
That is the consistency you were missing. It was never the model.
What this looks like when it works
Our client dossier is my favourite example.
When someone is going to Champagne, or Italy across three cities, they get a proper document. A summary of the trip. Their flights. The thinking behind why the trip looks the way it does. An hour by hour. And a real write up of every destination, so they know where they are going before they land.
Apps are great. But people want paper again. There is something about a printed thing you can read on the plane, or at the kitchen table with a glass of champagne, that a notification has never managed. 🥂
That document used to take my team about ten hours. It now takes about ten minutes.
Not because we found a magic tool. Because the process was written down, the folder had the right things in it, and the connectors meant it already knew what the client had told us on the call.
Week four: the three things I would build first
Now you get to do the thing everybody wanted to do in week one. Build these in this order.
The email auto draft
Ask it to schedule an automation that drafts responses to your inbox. Not sends. Drafts. You still read and press the button, but you are editing instead of starting from nothing.
The inbox dashboard
Have it check everything coming in and group it by project, so you open your day and immediately see what matters. This is the one that stops things falling through.
Your brochures
If you do not have professional level marketing brochures, you can have them by Friday. Share links from your website or your Drive that list your products, and have it build them.
One rule on that last one. Show it a design you like, because it is not psychic. Do not lift someone's design work. Go to Creative Market, buy a design you love, and hand that over as the reference. Pay the designer. It is thirty dollars and it is the right thing to do.
What I would tell you not to do
Do not let it run without you.
AI is not equipped to be autonomous, because it does not know what is in your head. It works with you. The second you leave it alone with something that matters, you get confident, well formatted nonsense.
Do not skip the QA.
Mine has a standing instruction to have an agent check the answer before it ever reaches me. Yes, it takes a little longer. It also means I am not the last line of defence at eleven at night.
Do not accept the first answer.
Ask what kind of expert does this work, then tell it to channel the top two and give you both, including where they disagree.
Do not ask for a deliverable without showing it one.
If you want a document, show it a document you like. If you want a deck, show it your last good deck.
The two things I never let it do
It never touches a payment.
Not a card, not an invoice, not a refund. Money is a human decision and it stays one.
It never deletes anything.
It can draft, sort, rename, summarise and suggest. It does not get delete permissions, and if a tool insists on them, that tool does not get my files.
Everything else is negotiable. Those two are not.
Somebody who actually did this
One of my VIP clients is building Black Wall Street Travel, and she launches next month.
I watched her team go from getting started, to posting on social, to making scroll stopping content with audio and video. I sent the team a message asking how they made it, because it was beautiful.
They said, “Claude!”
AI is the revolution that can take your business much further on a much smaller budget.
When I started, a website was ten thousand dollars over time. Email marketing was six thousand. StoryBranding was three thousand. And the list goes on and on.
Today you can do all of that yourself, with AI, for under fifteen hundred dollars.
I have said for a long time that ideas come to die with entrepreneurs who do not have access to information or the money to birth the idea. AI changes that. That is the part I actually care about.
If you only do one thing this week
Open your settings and tell it to play devil's advocate on everything.
That is it. One line. It takes ninety seconds and it will change every answer you get from here on out.
Then, when you have another hour, write the rest of your personalization instructions and start one chat in voice mode where you let it interview you before it answers.
Stop asking your AI for permission. Start asking it for the argument. 🍒
Now tell me what you are stuck on. Sign up just above and reply to the first email you get from me, or write to me directly, and I will answer you personally. I said I would answer every email and I meant it.
I am co-hosting an AI Build Day on Saturday, August 29 in the DC area, where we do all four of these weeks together in one day. Early registration is $397 through Friday. Details are here. And if you would rather do it on your own with this piece and four hours a week, that works too, and I mean that.
See the AI Build DayWarmly,
Claire
P.S. Still on the 2018 Lanzerac Pinotage. Still with the salted butter peanut butter and jelly sandwich. I am not taking questions at this time. What are you drinking this week?
Everything referenced above, in one place
Resources
- My AI Starter KitThe exact personalization instructions I use, free
- travelaiuniversity.com/starter-kit
- The V.O.I.C.E. frameworkHow to let AI interview you
- The capstone case study
- Download Claude CoworkSo it lives on your computer, not just a browser tab · the first 3 people get a week free
- Download Claude
- My weekly Travel AI Setup WebinarWe walk through this setup live
- Save your seat
- Creative MarketBuy the design, do not lift it
- creativemarket.com
- Black Wall Street TravelLaunching next month
- @blackwallstreettravel
- The AI Build DaySaturday, August 29 · DC area
- clairebsoares.com/aiworkshop
- Claire's newsletterThe next piece, straight to your inbox
- Join the list
- Join Travel AI UniversityThe community where we do all of this together
- skool.com/clairebsoares
Frequently Asked Questions
The questions I get most
- Why does my AI agree with everything I say?
- Most AI assistants agree with almost everything you say by default, because nothing in the setup tells them not to. The fix is a single line in your personalization settings: tell it to play devil's advocate on everything, every answer, whether you asked for it or not. It stops flattering you and starts forcing you to make a decision.
- Why is my AI inconsistent and inaccurate, brilliant one day and wrong the next?
- Your AI is not inconsistent and it is not inaccurate; your setup is. The same question on a different day with different context produces a different answer. Accuracy in particular is not a feature you get, it is an instruction you give. Nobody tells the model to verify anything, then everybody is shocked when it invents something confidently. Open the personalization section in Claude or ChatGPT and write standing instructions: opinionated answers, pros and cons, brief and to the point, clarifying questions before answering, fact-check like a New York Times fact checker, and quote your sources. Those apply to every chat, forever, instead of one at a time.
- Where do I actually start with AI if I have four hours a week?
- Four weeks, four hours each. Week one: write your personalization instructions and connect AI to the two or three places your business actually lives (usually email, calendar and file storage). Week two: pick the one job you resent and build it a folder of your real work. Week three: use voice mode and let the AI interview you before it produces anything, then write the process steps down. Week four: build the email auto-draft, the inbox dashboard, and your brochures.
- What is the difference between a butter knife and a steak knife setup?
- A butter knife is AI drafting your email with no knowledge of your policies, products, terms or how you actually speak to a nervous client. Technically possible, slow and undignified. A steak knife is the same tool after you have given it your trips, services, terms and conditions and real sent emails. The gap between the two experiences is not the software. It is about forty minutes of setup.
- What should I never let AI do?
- Two things are non-negotiable. It never touches a payment. No card, no invoice, no refund, because money is a human decision. And it never deletes anything; it can draft, sort, rename, summarise and suggest, but it does not get delete permissions. If a tool insists on them, that tool does not get your files.
- What is a comedy alert?
- It is a standing instruction I add to my personalization settings: if the AI makes anything up or cannot verify something, it has to say “comedy alert” and then list exactly what it invented. The model tells on itself at the bottom of every answer, which catches more errors than re-reading a paragraph and wondering whether it is true.
More stories like this
The next one, straight to you
No cost, no funnel. The same as this one. What is actually working, what broke, and the setup behind it, whenever I have something worth sending.
Unsubscribe any time. I do not share your email.
Want the setup, not the story?
Two things do most of the work in this piece, and both are free.