@jasoncooperson Comment "COUNCIL" and I'll send you the skill for free 👇
Stanford just proved Claude agrees with you 49% more than a human would. Translation: your AI is the world's greatest yes-man — hyping up your dumbest ideas instead of telling you the truth.
The fix comes from Andrej Karpathy (founding member at OpenAI, just left to join Anthropic). He built something called the LLM Council. I ripped his GitHub repo, patched it into a clean little skill, and now it runs with one command.
Type /llm-council and it spawns 5 advisor agents:
– a contrarian who only hunts for what'll fail
– a first-principles thinker who rips apart your assumptions
– an expansionist who finds the upside you're missing
– an outsider who knows nothing about your industry
– an executor who only cares what you do Monday morning
Then they blind-review each other's answers without knowing who wrote what, and a chairman agent makes the final call + next steps.
No more echo chamber. Real pushback on tap.
Comment "COUNCIL" and it's yours, free 👇
AI problem and council solutionfrom @jasoncooperson
[01]
Overview
The creator explains a significant issue with the AI model Claude, citing a Stanford study that found it to be a sycophant. He then introduces a solution created by Andrej Karpathy called the 'LLM Council,' which he has packaged into a free tool. The creator proceeds to break down how the tool works, listing and explaining five different AI advisor agents (Contrarian, First Principles Thinker, etc.) that provide diverse perspectives, followed by a blind review process and a final verdict from a chairman agent. The video concludes with a direct call-to-action, instructing viewers to comment 'COUNCIL' to receive the free tool and a guide.
A new study from Stanford just came out that proves Claude agrees with you 49% more than a human would.
Visual hook · 5%
A clear, direct-to-camera shot of the speaker in a well-lit office setting. He uses active hand gestures to immediately engage the viewer and convey energy.
[06]
Hook template
Steal this for your own video
[Credible source] just exposed a huge [subject] problem
— Harvard just exposed a huge sleep supplement problem
— The FDA just exposed a huge protein powder problem
— McKinsey just exposed a huge remote work problem
[07]
Hook and payoff
The promise the opener makes, and how the video pays it back
Hook
The Surprising Flaw: Present a startling, data-backed problem with a widely used tool or concept, immediately creating doubt and a need for a solution.
Payoff
The Value-Exchange CTA: Offer the creator's simplified version of the expert solution for free in exchange for a specific engagement action (e.g., a comment), reinforcing the value provided and driving interaction.
[08]
Narrative framework
The structural pattern this video runs on
Framework
The Expert's Fix
Identify a common tool's hidden flaw, present a solution created by a recognized authority, and offer a simplified, accessible version of that solution in exchange for engagement.
1
The Problem Amplification
Translate the data-backed problem into a relatable, negative real-world consequence for the viewer, making the threat feel personal.
2
The Authority Hand-off
Introduce the solution by attributing it to a highly credible, external authority figure, instantly validating its effectiveness and bypassing the need for the creator to build trust from scratch.
3
The Creator's Bridge
Position the creator as the one who has taken the expert's complex solution and made it simple and accessible for the audience, adding a layer of personal value.
4
The Systematic Breakdown
Detail the components of the solution in a list format, explaining the function of each part to demonstrate its robustness and comprehensiveness.
5
The Process Synthesis
Explain how the individual components work together in a multi-stage process to produce a final, superior, and vetted outcome.
Themes
Critical Thinking
AI Limitations
Problem-Solving
Open Source Tools
Techniques
Problem/Solution Framing
Expert Endorsement
Listicle
Call-to-Action
Trends
AI Tools
Creator Education
Post caption
Comment "COUNCIL" and I'll send you the skill for free 👇
Stanford just proved Claude agrees with you 49% more than a human would. Translation: your AI is the world's greatest yes-man — hyping up your dumbest ideas instead of telling you the truth.
The fix comes from Andrej Karpathy (founding member at OpenAI, just left to join Anthropic). He built something called the LLM Council. I ripped his GitHub repo, patched it into a clean little skill, and now it runs with one command.
Type /llm-council and it spawns 5 advisor agents:
– a contrarian who only hunts for what'll fail
– a first-principles thinker who rips apart your assumptions
– an expansionist who finds the upside you're missing
– an outsider who knows nothing about your industry
– an executor who only cares what you do Monday morning
Then they blind-review each other's answers without knowing who wrote what, and a chairman agent makes the final call + next steps.
No more echo chamber. Real pushback on tap.
Comment "COUNCIL" and it's yours, free 👇
Text overlays
title · beginning
Stanford just exposed a huge 💥 Claude problem
informational · beginning
Compared to human responses, all of the AIs affirmed the user's position more frequently. In the general advice and Reddit-based prompts, the models on average endorsed the user 49% more often than humans. Even when responding to the harmful prompts, the models endorsed the problematic behavior 47% of the time.
contextual · beginning
YES man
informational · middle
5 ADVISORS:
informational · middle
1. Contrarian
informational · middle
2. First Principles Thinker
informational · middle
3. Expansionist
informational · middle
4. Outsider
informational · middle
5. Executor
call-to-action · end
"COUNCIL"
Transcript
A new study from Stanford just came out that proves Claude agrees with you 49% more than a human would. This basically means that anytime you ask Claude for a big business decision or a question, it's basically just going to be the world's greatest yes rank, even when your idea is dumb as Here's how to fix this with one single prompt. The fix actually comes from the Good Andrej Karpathy. Karpathy was a founding member at OpenAI who recently left to work for Anthropic, and he built something called the LLM Council. I already ripped his GitHub repo, I took a look at all the files, and I built and packaged it up into a nice little skill that I'm going to give to you guys right now. Basically all you do is type llm council and it's going to spawn in five different advisor agents. The first agent is a contrarian who's only gonna look at what will fail. The second agent is a first principles thinker who will rip apart all of your assumptions. The third is an expansionist who focuses on the upside that you're missing. The fourth agent is an outsider who knows nothing about your industry. And the last agent is an executor who really only cares about what you're gonna do on Monday morning. Then each advisor reviews the other's answer without knowing who wrote which one. Lastly, a chairman agent comes in, reviews everything, gives you a final call and next steps. If you comment council down below, I will send you the skill that I took hours to make for completely free and a breakdown on how you can run a workflow that's actually faithful to the original design by Carpathi by spawning in these different providers. I hope you enjoy the free sauce. Peace out.
