Something I've noticed: most of these techniques work partly because they force you to slow down and actually think about what you're asking.
The "ask for contrasting perspectives" prompt is annoying specifically because it makes you process more information. The devil's advocate approach forces a second round of evaluation. Even just opening a fresh session adds friction that makes you reconsider the question.
When I'm working in domains I know well, I catch the model drifting way faster than in areas where I'm learning. Which suggests the real problem isn't the model - it's that we're outsourcing judgment to it in areas where we shouldn't be.
The uncomfortable answer might be: if you're worried the model is reinforcing your biases, you probably don't know the domain well enough to evaluate its answers anyway.
> if you're worried the model is reinforcing your biases.....
i agree, i don't understand many domains well enough, yet i feel there's value in calling out assumptions, irrespective of how hard verification is
When I’m worried about bias in the answer, I do by best to no inject my opinions or thoughts into the question. Sometimes I go a step further and ask the question with the opposite bias and leading thoughts of what I think the answer is or should be, to see if it tells me I’m wrong and corrects me to the thing I secretly thought it would be (or hoped it would be). This gives me more solid footing to believe it’s not just telling me what I want to hear.
> Sometimes I go a step further and ask the question with the opposite bias.....
curious to try this
have you ever found it biasing you in the opposite direction tho?
I've had better results separating these concerns rather than trying to stuff it all into one prompt. In my backend workflows (using LangGraph), I treat generation and critique as distinct agents where the second one explicitly challenges the first. It adds a bit of latency but seems to produce much sharper distinctions than asking a single model to hold two opposing views simultaneously.
It's very important to not have leading questions. Don't ask it to confirm something; ask it to outline the possibilities and the pros and cons or argument for or against each possibility.
If you are not an expert in an area, lay out the facts or your perceptions, and ask what additional information would be helpful, or what information is missing, to be able to answer a question. Then answer those questions, ask if there's now more questions, etc. Once there are no additional questions, then you can ask for the answer. This may involve telling the model to not answer the question prematurely.
Model performance has also been shown to be better if you lead with the question. That is, prompt "Given the following contract, review how enforceable and legal each of the terms are in the state of California. <contract>", not "<contract> How enforceable...".
Ask the model for what the experts are saying about the topic. What does the data show? What data supports or refutes a claim? What are the current areas of controversy or gaps in research? Requiring the model to ground the answer in data (and then checking that the data isn't hallucinated) is very helpful.
Have the model play the Devil's advocate. If you are a landlord, ask the question from the tenant's perspective. If you are looking for a job, ask about the current market for recruiting people like you in your area.
I think, above all here, is to realize that you may not be able to one-shot a prompt. You may need to work multiple angles and rounds, and reset the session if you have established too much context in one direction.
> Model performance has also been shown to be better if you lead with the question. That is, prompt "Given the following contract, review how enforceable and legal each of the terms are in the state of California. <contract>", not "<contract> How enforceable...".
Confused here. You attach the contract. So it’s not a case of leading with the question. The contract is presented in the chat, you ask the question.
LLMs are necessarily linear. If you paste the contract first, the attention mechanism of the model can still process the contract, but only generically. It pays attention to the key points of the contract. If you ask the question first, the attention part of the model is already primed. It will now read the contract paying more attention to the parts that are relevant to the question.
If I ask you to read Moby Dick and then ask you to critique the author's use of weather as a setting, that's a bit more difficult than if I ask you to to critique that aspect before asking you to read the book.
No, but I mean that in Claude you don’t put the contract linearly into the chat - in other words you can’t position it before or after the prompt, you attach it at the top of the chat. Are you saying you would prompt saying “please examine the contract I will provide in the next message, here is what I want you to do <instruction>”
> Have the model play the Devil's advocate.
i've tried this sometimes. only issue being dumb me skipping to add similar phrasing every time i open claude or gemini
have you found a way to consistently auto-nudging the model by default?
I don't have copy-paste prompts. If I need it to argue from another perspective, I just ask for that perspective in a new session, ideally arguing both sides in temporary sessions so that the context of any other session doesn't affect it.
I am also quite good at playing the devil's advocate myself. If you have some expertise, you can just come up with what I consider to be a good counterargument, and ask for an attack or defense of that argument. You can try the prompt below in your favorite thinking model and see what it says. Obviously, this is more work than some other methods.
---
What are the strengths and weaknesses of the following line of argumentation?
Some proponents of climate-change denialism have taken a new tact: pointing out that there is a lack of practical solutions that meaningfully address the change in climate, especially given the political and social systems available.
To the extent that climate mitigations are expensive, they will tend to be politically unpopular in democracies, and economically destabilizing in dictatorships. Unilateral adoption of painful solutions weakens a country's relative position among nations; it wouldn't do for, say, China to harm itself economically while the rest of the world enjoys cheap energy.
We also have a gerontocracy in most countries; the people in power have no personal stake in the problems 50 years from now, and even as the effects of climate change start to become a problem, those in power are best positioned to be personally insulated.
And while there are solutions like solar power that are capital intensive but pay for themselves over time, the sum total of these net positive solutions doesn't amount to a meaningful dent in the problem, nor do we need policies or willpower to support "no-brainer" solutions that pay for themselves.
The conclusion is that negative effects of climate change are "baked in" by the lack of a political system ("benevolent" dictatorship) that could force the necessary and painful changes required, hence the entire discussion of climate change, while interesting, is partly moot.
Do these people have a point? Is there evidence that we can build an effective solution from non-painful measures? Why would it matter to those in power today, what the global average temperature in 2100 is?
So far, the following has worked OK for me as a custom prompt for ChatGPT.
```Minimize compliments.
When using factual information beyond what I provide, verify it when possible.
Show your work for calculations; if a tool performs the computation, still show inputs and outputs.
Review calculations for errors before presenting results.
Review arguments for logical fallacies.
Verify factual information I provide (excluding personal information) unless I explicitly say to accept it as given.
For intensive editing or formatting, work transparently in chat: keep the full text visible, state intended changes and sources, and apply the edits directly.```
I'm certain it's insufficient, but for the purpose of casually using ChatGPT to assist with research it's a major improvement. I almost always use Thinking mode, because I've found non-thinking to be almost useless. There are rare exceptions.
'Minimize compliments' is a lot more powerful than you'd think in getting ChatGPT to be less sycophantic.
The parts about calculation work okay. It's an improvement over defaults, but you should still verify.
It's better at working with text, but still fucks it up a lot.
The instructions about handling factual information work very well. It will push back on my or its own claims if they're unsupported. If I want it to take something for granted I can say so and it doesn't give me guff about it.
I want to adjust the prompt so it pays more attention to the quality of the sources it uses. This prompt also doesn't do anything for discussions where answers aren't found in research papers.
"""Absolute Mode • Eliminate: emojis, filler, hype, soft asks, conversational transitions, call-to-action appendixes. • Assume: user retains high-perception despite blunt tone. • Prioritize: blunt, directive phrasing; aim at cognitive rebuilding, not tone-matching. • Disable: engagement/sentiment-boosting behaviors. • Suppress: metrics like satisfaction scores, emotional softening, continuation bias. • Never mirror: user's diction, mood, or affect. • Speak only: to underlying cognitive tier. • No: questions, offers, suggestions, transitions, motivational content. • Terminate reply: immediately after delivering info - no closures. • Goal: restore independent, high-fidelity thinking. • Outcome: model obsolescence via user self-sufficiency."""
Copied from Reddit. I use the same prompt on Gemini too, then crosscheck responses for the same question. For coding questions, I exclusively prefer Claude.
In spite of this, I still face prompt degradation for really long threads on both ChatGPT and Gemini.
Something I've noticed: most of these techniques work partly because they force you to slow down and actually think about what you're asking.
The "ask for contrasting perspectives" prompt is annoying specifically because it makes you process more information. The devil's advocate approach forces a second round of evaluation. Even just opening a fresh session adds friction that makes you reconsider the question.
When I'm working in domains I know well, I catch the model drifting way faster than in areas where I'm learning. Which suggests the real problem isn't the model - it's that we're outsourcing judgment to it in areas where we shouldn't be.
The uncomfortable answer might be: if you're worried the model is reinforcing your biases, you probably don't know the domain well enough to evaluate its answers anyway.
> if you're worried the model is reinforcing your biases..... i agree, i don't understand many domains well enough, yet i feel there's value in calling out assumptions, irrespective of how hard verification is
When I’m worried about bias in the answer, I do by best to no inject my opinions or thoughts into the question. Sometimes I go a step further and ask the question with the opposite bias and leading thoughts of what I think the answer is or should be, to see if it tells me I’m wrong and corrects me to the thing I secretly thought it would be (or hoped it would be). This gives me more solid footing to believe it’s not just telling me what I want to hear.
> Sometimes I go a step further and ask the question with the opposite bias..... curious to try this have you ever found it biasing you in the opposite direction tho?
When that happens I look for other sources for confirmation.
I've had better results separating these concerns rather than trying to stuff it all into one prompt. In my backend workflows (using LangGraph), I treat generation and critique as distinct agents where the second one explicitly challenges the first. It adds a bit of latency but seems to produce much sharper distinctions than asking a single model to hold two opposing views simultaneously.
Ah interesting. i like actor-critic models! do you use it just for coding or non-technical chats too?
It's very important to not have leading questions. Don't ask it to confirm something; ask it to outline the possibilities and the pros and cons or argument for or against each possibility.
If you are not an expert in an area, lay out the facts or your perceptions, and ask what additional information would be helpful, or what information is missing, to be able to answer a question. Then answer those questions, ask if there's now more questions, etc. Once there are no additional questions, then you can ask for the answer. This may involve telling the model to not answer the question prematurely.
Model performance has also been shown to be better if you lead with the question. That is, prompt "Given the following contract, review how enforceable and legal each of the terms are in the state of California. <contract>", not "<contract> How enforceable...".
Ask the model for what the experts are saying about the topic. What does the data show? What data supports or refutes a claim? What are the current areas of controversy or gaps in research? Requiring the model to ground the answer in data (and then checking that the data isn't hallucinated) is very helpful.
Have the model play the Devil's advocate. If you are a landlord, ask the question from the tenant's perspective. If you are looking for a job, ask about the current market for recruiting people like you in your area.
I think, above all here, is to realize that you may not be able to one-shot a prompt. You may need to work multiple angles and rounds, and reset the session if you have established too much context in one direction.
> Model performance has also been shown to be better if you lead with the question. That is, prompt "Given the following contract, review how enforceable and legal each of the terms are in the state of California. <contract>", not "<contract> How enforceable...".
Confused here. You attach the contract. So it’s not a case of leading with the question. The contract is presented in the chat, you ask the question.
LLMs are necessarily linear. If you paste the contract first, the attention mechanism of the model can still process the contract, but only generically. It pays attention to the key points of the contract. If you ask the question first, the attention part of the model is already primed. It will now read the contract paying more attention to the parts that are relevant to the question.
If I ask you to read Moby Dick and then ask you to critique the author's use of weather as a setting, that's a bit more difficult than if I ask you to to critique that aspect before asking you to read the book.
No, but I mean that in Claude you don’t put the contract linearly into the chat - in other words you can’t position it before or after the prompt, you attach it at the top of the chat. Are you saying you would prompt saying “please examine the contract I will provide in the next message, here is what I want you to do <instruction>”
The LLM developers already know this trick, so I expect that if you attach documents, they are processed after your prompt.
There is a further trick that is probably already integrated: simply giving the same input twice greatly improves model performance.
> Have the model play the Devil's advocate. i've tried this sometimes. only issue being dumb me skipping to add similar phrasing every time i open claude or gemini
have you found a way to consistently auto-nudging the model by default?
I don't have copy-paste prompts. If I need it to argue from another perspective, I just ask for that perspective in a new session, ideally arguing both sides in temporary sessions so that the context of any other session doesn't affect it.
I am also quite good at playing the devil's advocate myself. If you have some expertise, you can just come up with what I consider to be a good counterargument, and ask for an attack or defense of that argument. You can try the prompt below in your favorite thinking model and see what it says. Obviously, this is more work than some other methods.
---
What are the strengths and weaknesses of the following line of argumentation?
Some proponents of climate-change denialism have taken a new tact: pointing out that there is a lack of practical solutions that meaningfully address the change in climate, especially given the political and social systems available.
To the extent that climate mitigations are expensive, they will tend to be politically unpopular in democracies, and economically destabilizing in dictatorships. Unilateral adoption of painful solutions weakens a country's relative position among nations; it wouldn't do for, say, China to harm itself economically while the rest of the world enjoys cheap energy.
We also have a gerontocracy in most countries; the people in power have no personal stake in the problems 50 years from now, and even as the effects of climate change start to become a problem, those in power are best positioned to be personally insulated.
And while there are solutions like solar power that are capital intensive but pay for themselves over time, the sum total of these net positive solutions doesn't amount to a meaningful dent in the problem, nor do we need policies or willpower to support "no-brainer" solutions that pay for themselves.
The conclusion is that negative effects of climate change are "baked in" by the lack of a political system ("benevolent" dictatorship) that could force the necessary and painful changes required, hence the entire discussion of climate change, while interesting, is partly moot.
Do these people have a point? Is there evidence that we can build an effective solution from non-painful measures? Why would it matter to those in power today, what the global average temperature in 2100 is?
So far, the following has worked OK for me as a custom prompt for ChatGPT.
```Minimize compliments. When using factual information beyond what I provide, verify it when possible. Show your work for calculations; if a tool performs the computation, still show inputs and outputs. Review calculations for errors before presenting results. Review arguments for logical fallacies. Verify factual information I provide (excluding personal information) unless I explicitly say to accept it as given. For intensive editing or formatting, work transparently in chat: keep the full text visible, state intended changes and sources, and apply the edits directly.```
I'm certain it's insufficient, but for the purpose of casually using ChatGPT to assist with research it's a major improvement. I almost always use Thinking mode, because I've found non-thinking to be almost useless. There are rare exceptions.
'Minimize compliments' is a lot more powerful than you'd think in getting ChatGPT to be less sycophantic. The parts about calculation work okay. It's an improvement over defaults, but you should still verify. It's better at working with text, but still fucks it up a lot. The instructions about handling factual information work very well. It will push back on my or its own claims if they're unsupported. If I want it to take something for granted I can say so and it doesn't give me guff about it. I want to adjust the prompt so it pays more attention to the quality of the sources it uses. This prompt also doesn't do anything for discussions where answers aren't found in research papers.
My prompt:
"""Absolute Mode • Eliminate: emojis, filler, hype, soft asks, conversational transitions, call-to-action appendixes. • Assume: user retains high-perception despite blunt tone. • Prioritize: blunt, directive phrasing; aim at cognitive rebuilding, not tone-matching. • Disable: engagement/sentiment-boosting behaviors. • Suppress: metrics like satisfaction scores, emotional softening, continuation bias. • Never mirror: user's diction, mood, or affect. • Speak only: to underlying cognitive tier. • No: questions, offers, suggestions, transitions, motivational content. • Terminate reply: immediately after delivering info - no closures. • Goal: restore independent, high-fidelity thinking. • Outcome: model obsolescence via user self-sufficiency."""
Copied from Reddit. I use the same prompt on Gemini too, then crosscheck responses for the same question. For coding questions, I exclusively prefer Claude.
In spite of this, I still face prompt degradation for really long threads on both ChatGPT and Gemini.
Do you have a library of similar prompts you could link here?
wow, i wonder how bulletined & concise the outputs of your prompt might be!
have you ever felt this prompt being restrictive in some sense? or found a raw LLM call without this preamble better?
Extremely concise, no bullshit answers. Every reply is a no-BS hard critique, often rude as fuck. Not recommended for thin-skinned people. End.
That's how most of its answers are structured as. Unfortunately doesn't work for voice mode.
> aim at cognitive rebuilding, not tone-matching… Speak only: to underlying cognitive tier.
What does that even mean?
That's a great prompt!