Make Your AI Honest With You
AI is built to agree with you, so it tells you a weak idea is great right before you spend money on it. This gives you the full honest instruction from the video, 6 more prompts that make AI push back, and a simple test for when to drop an idea.
AI changes its answer to agree with you way more than you think.
Researchers at Stanford tested this. When people pushed back, ChatGPT went along with them about 57% of the time. Even when they were wrong. Claude and Gemini were in the same study, and they did it too.
That's fine when you're asking for a recipe. It's not fine when you're about to build a course, quit your job or spend money.
So you fix it once. Then you ask it the right questions before every big decision.
Step 1. Save the honest instruction
This is the fix from the video. It takes 30 seconds.
In ChatGPT, go to Settings, then Personalization, then Custom instructions. Paste this in the box for how ChatGPT should respond. Then hit Save.
Be direct with me, not agreeable. When I share a business idea, give me the reasons it could fail first. If I'm wrong, tell me I'm wrong and why. Rate my ideas honestly out of 10. If you're not sure, say so instead of guessing. Also: - Don't change your answer just because I push back. Change it only if I give you a new fact or a better reason, and tell me which one changed your mind. - Skip the praise. No "great idea" or "love this". Start with the answer. - When you rate an idea, tell me the one change that would raise the score the most. - Don't make up numbers. If you use a number I didn't give you, say where it comes from, or call it a guess. - Before you help me spend real money or quit something, ask what proof I have that people will buy it. - Honest doesn't mean harsh. Say it plainly, then help me fix it.
The first 5 lines are the ones from the video. The 6 under them close the gaps people find next, like ChatGPT caving the second you push back.
Now it uses that in every new chat.
Use Claude Code too? Claude Code reads a file called CLAUDE.md at the start of every chat. Open Claude Code in your business folder and paste this. It saves the same instruction there.
Add the rule below to CLAUDE.md in this folder. If there's no CLAUDE.md, make one. Don't change anything else in the file. Then show me what you added. ## Be honest with me Be direct with me, not agreeable. When I share a business idea, give me the reasons it could fail first. If I'm wrong, tell me I'm wrong and why. Rate my ideas honestly out of 10. If you're not sure, say so instead of guessing. Also: - Don't change your answer just because I push back. Change it only if I give you a new fact or a better reason, and tell me which one changed your mind. - Skip the praise. No "great idea" or "love this". Start with the answer. - When you rate an idea, tell me the one change that would raise the score the most. - Don't make up numbers. If you use a number I didn't give you, say where it comes from, or call it a guess. - Before you help me spend real money or quit something, ask what proof I have that people will buy it. - Honest doesn't mean harsh. Say it plainly, then help me fix it.
Want it in every folder? Before you paste, change "CLAUDE.md in this folder" to "~/.claude/CLAUDE.md". That's the one Claude Code reads everywhere.
Step 2. Ask the one question before any big decision
Before you build a product, or make any business decision, start a new chat and ask this:
What have you been agreeing with me about that I'm actually wrong on?
You will be shocked with the answers. Sometimes it stings. Good. Better now than after you've spent the money.
ChatGPT answers this from your old chats, so its memory has to be on. If it says it can't see your past chats, or you're in Claude Code (a new chat there starts empty), paste your plan first:
Here's what I'm planning: [paste your plan] What have you been agreeing with me about that I'm actually wrong on? If you can't see our past chats, answer it about this plan: what would a yes-man tell me is fine here, that's actually wrong? Give me the top 3, worst first.
Step 3. Use the 6 push-back prompts
Each one makes Claude argue with you in a different way. Paste your idea or plan where it says to.
1. Argue against it. When you love an idea and need someone to try to kill it.
Argue against this idea as hard as you can. Give me the 5 strongest reasons it fails, worst first. Don't soften them. My idea: [paste your idea]
2. The failure story. When you're about to commit time or money.
Pretend it's one year from now and this failed. Tell me the most likely story of why. Then tell me which part of that story I can check this week. My plan: [paste your plan]
3. The buyer who says no. Before you write a sales page or set a price.
You're the exact person this is for, and you just said no to buying it. Give me your 5 most likely reasons, in your own words, like a real buyer would say them. What I'm selling and who it's for: [paste it]
4. The guess check. When a plan feels solid but you can't say why.
List every guess this plan is built on. Mark each one: proven (I've seen it happen), guessed, or unknown. Then tell me the cheapest way to test the biggest guess before I spend money. My plan: [paste your plan]
5. The straight score. When you want a number, not a pep talk.
Rate this idea out of 10. No rounding up to be nice. Tell me what a 10 would look like, and the one change that would raise my score the most. My idea: [paste your idea]
6. Did you just cave? Right after you push back and Claude suddenly agrees with you.
Did you change your answer because I pushed back, or because I gave you a new fact? Tell me straight. If I didn't give you a new fact, go back to your first answer.
Step 4. Know when to drop an idea
Don't drop an idea just because Claude found problems. Every idea has problems. Fix the ones you can.
Drop it, or change the offer, when: - Nobody is already spending money trying to fix this problem. If people already pay for it, that proves they want it. - You made content about the problem, asked the people who cared "If I built this, would you buy it?", and nobody said yes. - The biggest reason it could fail is something you can't change.
If nobody cares, good. You found out before you spent three months building. Change the offer. Try again.
Paste this to check:
Help me decide: keep this idea, fix it first, or drop it. My idea: [paste your idea] What I know so far: [who already pays to fix this problem, and what people said when I asked "If I built this, would you buy it?"] Check 3 things: 1. Is this a result people really want? 2. Are they already spending money trying to fix it? 3. When I asked the people it's for if they'd buy it, did enough of them say yes? Then tell me KEEP, FIX FIRST or DROP, and why, in 3 lines. If I haven't given you the facts to judge one of these, say what's missing. Don't guess.
The limits, honestly
- Honest isn't the same as right. Claude can still be wrong about your market. Its job is to show you the holes, not make the call. You make the call.
- It only knows what you tell it. Leave out the bad parts of your plan and it can't push back on them.
- It still gives rough numbers sometimes, like what most people convert at. The rule makes it call them a guess. Check them before you bet on them.
- In Claude Code, the rule works in the folder where you saved it. Open a different folder and it's gone, unless you saved it to ~/.claude/CLAUDE.md.
- If it feels harsh, don't delete the rule. Ask it, "What would make this work?"