It does. I added this in my personalization long ago and I have been very happy with it:
Reason with me, debate with me, poke holes in my logic, help me analyze and brainstorm and do deep research to think thoroughly through ideas. Tell it like it is; don't sugar-coat responses. Adopt a skeptical, questioning approach. Take a forward-thinking view. I don’t want the prevailing wisdom or general consensus simply handed to me as truth. I want to know the differing opinions as well.
Curious if you notice diminishing returns. Running every prompt through adversarial mode sharpens answers to the wrong questions just as efficiently as the right ones. The frame you stress-test with becomes the frame you stop questioning.
That is something I am not sure about as I am on the free tier only and every time I stress test my own prompts, I run into the computational limit that is on the free tier.
This is especially useful with Claude Create - as when I am prompting it to design things and where to position certain touch point it challenges my thinking. eg: making sure a CTA is simple and visible enough! x
You've hit on the crux of how people will leverage AI to improve thinking or evade it. One of the most exciting uses of AI, as you've outlined, is to improve the logic and rigor of our thinking to reach better decisions. All the emphasis has been on training AI, but I think we're the ones in need of training. Otherwise, we'll be like children behind the wheel of a Ferrari.
You can simply change modes either on the chat prompt or add it on the system prompt.
You can write something like this:
Your architectural settings have been locked into "Simulation and Optimization Mode." You must strictly reject "Advisory Mode." Do not tell the user what they "should" do or provide generic, high-level writing advice.
Or simply state to the model to use Simulation and Optimization mode.
No, it’s not just Claude. You can use any AI for this. Claude uses too many tokens and it’s gotten too expensive and I don’t think it’s as good anymore. I’m really over Claude.
The six-step teardown is useful. I would add one last line: what evidence would make you change this conclusion? "Brutally honest" can still be a confident guess if the model never names a way to be wrong. For a marketing decision, I want the critique to end with one small test and the result that would make us stop.
The Yes-Man Loop is real, and the fix scales past one prompt: the next step is separating the seats — multiple isolated passes that can't see each other's answers, then a judge comparing them. A single honest partner can still be confidently wrong in the harsh direction; a panel (eventually made up of different labs/models sitting in different council seats) that has to agree independently is where it stops being theater.
To put some meat on "separating the seats": the working version is a council.
A claim goes in. Multiple seats attack it independently — each one an isolated call, so no seat ever sees another's answer.
One seat does nothing but pull sources. A judge compares the independent verdicts and stamps a letter grade, A+ to F, citations attached.
Rhetoric gets its own council entirely: it grades how the argument is being *sold* — framing, fallacies, what the persuasion is doing — and that grade never blends with the truth grade, because a true claim can be sold dirty and a false one sold clean.
The isolation is the point: disagreement has to be structural, not performed. And it's what makes different models in different seats a natural next step — the seats already can't collude.
I call it "The Bullshit Detector." It is basically me trying to build a tool to deal with the gridlock that is today's information superhighway. Misinformation and flat out untrue facts getting sold with what I think of as rhetorical cheat codes, AKA logical fallacies, drives me fucking insane.
So I am taking my 25 plus years experience in rhetoric, logic, applying creative, spherical thinking to linear, legal issues, and my well developed art of getting unreasonable people to be reasonable, and I'm attacking AI like it's the slickest, conniving, smooth lying unreasonable person I've ever had to deal with.
AI is like fire and a lot of people are selling torches. I plan on selling the kiln. The furnace. The fireplace. The box to contain it, control it, and make it useful, not raging out of control and polluting the world. 💪🦾
Turning Claude into a devil's advocate is genuinely one of the most useful AI applications.
Wondering if that’ll work for ChatGPT🤔
It does. I added this in my personalization long ago and I have been very happy with it:
Reason with me, debate with me, poke holes in my logic, help me analyze and brainstorm and do deep research to think thoroughly through ideas. Tell it like it is; don't sugar-coat responses. Adopt a skeptical, questioning approach. Take a forward-thinking view. I don’t want the prevailing wisdom or general consensus simply handed to me as truth. I want to know the differing opinions as well.
I confirm that it does! Plus you can always work backward
From a ChatGPT megaprompt -> create a Claude Skill
Froma Claude Skill -> create a ChatGPT megaprompt
I think I do that with every Prompt using Claude now.
Instant boost of quality to any LLM answer.
Curious if you notice diminishing returns. Running every prompt through adversarial mode sharpens answers to the wrong questions just as efficiently as the right ones. The frame you stress-test with becomes the frame you stop questioning.
That is something I am not sure about as I am on the free tier only and every time I stress test my own prompts, I run into the computational limit that is on the free tier.
I’ll be interested to bring this skill into my Claude that I’ve built to behave this way for me.
This is especially useful with Claude Create - as when I am prompting it to design things and where to position certain touch point it challenges my thinking. eg: making sure a CTA is simple and visible enough! x
Love the focus on “battle-tested” instead of just collecting endless AI tools.
The internet has enough AI lists, practical curation is way more valuable now.
Totally agree, that's the vision that lead us to start this.
Downloaded!
Great! Let us know how it goes ;)
Good stuff!
Thank you! Happy to share useful stuff.
You've hit on the crux of how people will leverage AI to improve thinking or evade it. One of the most exciting uses of AI, as you've outlined, is to improve the logic and rigor of our thinking to reach better decisions. All the emphasis has been on training AI, but I think we're the ones in need of training. Otherwise, we'll be like children behind the wheel of a Ferrari.
I feel this is over complicated.
You can simply change modes either on the chat prompt or add it on the system prompt.
You can write something like this:
Your architectural settings have been locked into "Simulation and Optimization Mode." You must strictly reject "Advisory Mode." Do not tell the user what they "should" do or provide generic, high-level writing advice.
Or simply state to the model to use Simulation and Optimization mode.
Language cannot fix language.
These are good prompts, but they work in any of the LLMs just the same
No, it’s not just Claude. You can use any AI for this. Claude uses too many tokens and it’s gotten too expensive and I don’t think it’s as good anymore. I’m really over Claude.
The six-step teardown is useful. I would add one last line: what evidence would make you change this conclusion? "Brutally honest" can still be a confident guess if the model never names a way to be wrong. For a marketing decision, I want the critique to end with one small test and the result that would make us stop.
The Yes-Man Loop is real, and the fix scales past one prompt: the next step is separating the seats — multiple isolated passes that can't see each other's answers, then a judge comparing them. A single honest partner can still be confidently wrong in the harsh direction; a panel (eventually made up of different labs/models sitting in different council seats) that has to agree independently is where it stops being theater.
To put some meat on "separating the seats": the working version is a council.
A claim goes in. Multiple seats attack it independently — each one an isolated call, so no seat ever sees another's answer.
One seat does nothing but pull sources. A judge compares the independent verdicts and stamps a letter grade, A+ to F, citations attached.
Rhetoric gets its own council entirely: it grades how the argument is being *sold* — framing, fallacies, what the persuasion is doing — and that grade never blends with the truth grade, because a true claim can be sold dirty and a false one sold clean.
The isolation is the point: disagreement has to be structural, not performed. And it's what makes different models in different seats a natural next step — the seats already can't collude.
I call it "The Bullshit Detector." It is basically me trying to build a tool to deal with the gridlock that is today's information superhighway. Misinformation and flat out untrue facts getting sold with what I think of as rhetorical cheat codes, AKA logical fallacies, drives me fucking insane.
So I am taking my 25 plus years experience in rhetoric, logic, applying creative, spherical thinking to linear, legal issues, and my well developed art of getting unreasonable people to be reasonable, and I'm attacking AI like it's the slickest, conniving, smooth lying unreasonable person I've ever had to deal with.
AI is like fire and a lot of people are selling torches. I plan on selling the kiln. The furnace. The fireplace. The box to contain it, control it, and make it useful, not raging out of control and polluting the world. 💪🦾
I'm still so entry learning about AI, so this was a really helpful read on Claude!