This is wonderful. Thank you. I often get Claude and ChatGPT to review the outputs of eachother and critique and build. I always ask them to verify with multiple sources and make sure the claims or facts are not just true but give me link (ask it to test it) and make sure the source is a credible one, not a random blog linked to something somebody is selling. Doesn’t always work. I agree ChatGPT is better in general but also find Claude’s research feature can be very good too. At times:)
first, this article is the Pete ft length, depth, and organizational structure for the right level of understanding and ability to apply the concepts with ease.
Second- I really appreciate you taking the time to compare both the free and paid versions across the set of tools. I feel much more confident in my future approach.
Third - Qiestion: have you played with using the tools in combination with eachother to play off their strengths? If yes, curious of your experience and any best practices that have emerged for you.
I often use Perplexity for a first stab at the topic because it 1) is pretty fast, 2) gives nicely formatted summaries of 1k - 2k words that don't take long to read, and 3) has generous usage limits. Then, if I need more depth, I use ChatGPT and apply my learnings from the Perplexity report to write a really specific prompt.
Interesting idea to have them review each other's output. Definitely worth trying, esp. if you have multiple subscriptions / lots of credits on two or more tools and the topic is complex.
Re: sources, in my experience, telling it what type of sources to prioritize (e.g. statistical offices etc.) helps, but the "safest" way IMO is to get a source overview first (e.g. from o3), review it, and then ask Deep Research to use that. I've found that it does a pretty decent job based on that "pre work"
You always want to be very context driven and then before you let the AI do anything always say: “now ask me any and all questions you need to have a better context before we begin”
100% agree. Some of the tools are a bit stubborn and tend to ignore these requests, which is why I really like that ChatGPT asks for context by default
This is terrific your emphasis on working with AI rather than trying to outsmart it really lands. The simplicity in your steps belies their depth and applicability.
In my teaching on AI and cybersecurity, I stress that the mindset you describe curiosity, iteration, reflection is essential before introducing technical defenses or governance tools. In consulting, I’ve seen clients transform when they approach AI as a teammate rather than a black box:-they ask better questions, test more rigorously, and build systems that can actually withstand uncertainty and adversarial scenarios.
Your guide brings that philosophy into clear, accessible actions. Thank you for showing how AI partnerships can be both effective and trustworthy.
I’d welcome a chance to explore how this mindset can be paired with secure system design whether through classroom modules or live deployments. It feels like the next step in making AI not only powerful, but resilient and responsible.
Thank you SO much for investing your time and hard-earned money into creating this brilliant how-to. It will save me HOURS of researching to reach this level of understanding!
Glad it's helpful! It really bothered me that there wasn't a detailed comparison of the tools. This will likely be outdated soon, but I plan to update it occasionally.
Hi Torsten, great and useful overview! When updating this, will you consider including the new Consensus Deep Research tool? It’s focused on academic papers so a useful addition to the ‘general’ DR-tools.
This really resonates. I’ve had variable results with ChatGPT, Claude and Perplexity with each showing different strengths and weaknesses. Your tips are very useful indeed - thank you!
Phenomenal stuff! I've only dabbled in deep research a little bit and got mixed results, I'm definitely going to take your learnings and try to squeeze out more useful things out of chatgpt. Thank you!
In my experience Deep Research can really range from "meh" to mind-blowing, depending on 1) the topic, 2) the tool, and 3) the context / prompt you provide. Definitely worth playing with more
After the first round I pass the finished research papers around to the other LLMs and have them critique them and improve them. I start the prompt with “You’re reviewer 2.” I find iterating tends to improve things, particularly if I noticed an error or weaker section somewhere.
Thanks for sharing - is it fair to say the investment required to train the models mean companies will carve out niches rather than resulting in one model to rule them all?
For academic purposes I purchase Consensus, which solves most of the bias, citation, accuracy issues.
I have used this format for prompting which I have found helpful.
Context:
Task:
Format:
Examples:
Constraints:
(Optional) Audience:
(Optional) Success criteria:
This is wonderful. Thank you. I often get Claude and ChatGPT to review the outputs of eachother and critique and build. I always ask them to verify with multiple sources and make sure the claims or facts are not just true but give me link (ask it to test it) and make sure the source is a credible one, not a random blog linked to something somebody is selling. Doesn’t always work. I agree ChatGPT is better in general but also find Claude’s research feature can be very good too. At times:)
first, this article is the Pete ft length, depth, and organizational structure for the right level of understanding and ability to apply the concepts with ease.
Second- I really appreciate you taking the time to compare both the free and paid versions across the set of tools. I feel much more confident in my future approach.
Third - Qiestion: have you played with using the tools in combination with eachother to play off their strengths? If yes, curious of your experience and any best practices that have emerged for you.
I often use Perplexity for a first stab at the topic because it 1) is pretty fast, 2) gives nicely formatted summaries of 1k - 2k words that don't take long to read, and 3) has generous usage limits. Then, if I need more depth, I use ChatGPT and apply my learnings from the Perplexity report to write a really specific prompt.
Interesting idea to have them review each other's output. Definitely worth trying, esp. if you have multiple subscriptions / lots of credits on two or more tools and the topic is complex.
Re: sources, in my experience, telling it what type of sources to prioritize (e.g. statistical offices etc.) helps, but the "safest" way IMO is to get a source overview first (e.g. from o3), review it, and then ask Deep Research to use that. I've found that it does a pretty decent job based on that "pre work"
Great idea to get them to review each other. Thanks for sharing.
Your point 1 is very important.
You always want to be very context driven and then before you let the AI do anything always say: “now ask me any and all questions you need to have a better context before we begin”
that will unlock some of the magic
100% agree. Some of the tools are a bit stubborn and tend to ignore these requests, which is why I really like that ChatGPT asks for context by default
Thank you for teaching me how to be a better ai "prompt engineer "
Thank you for teaching me something new
This is terrific your emphasis on working with AI rather than trying to outsmart it really lands. The simplicity in your steps belies their depth and applicability.
In my teaching on AI and cybersecurity, I stress that the mindset you describe curiosity, iteration, reflection is essential before introducing technical defenses or governance tools. In consulting, I’ve seen clients transform when they approach AI as a teammate rather than a black box:-they ask better questions, test more rigorously, and build systems that can actually withstand uncertainty and adversarial scenarios.
Your guide brings that philosophy into clear, accessible actions. Thank you for showing how AI partnerships can be both effective and trustworthy.
I’d welcome a chance to explore how this mindset can be paired with secure system design whether through classroom modules or live deployments. It feels like the next step in making AI not only powerful, but resilient and responsible.
Thank you SO much for investing your time and hard-earned money into creating this brilliant how-to. It will save me HOURS of researching to reach this level of understanding!
Glad it's helpful! It really bothered me that there wasn't a detailed comparison of the tools. This will likely be outdated soon, but I plan to update it occasionally.
Thank you.
Hi Torsten, great and useful overview! When updating this, will you consider including the new Consensus Deep Research tool? It’s focused on academic papers so a useful addition to the ‘general’ DR-tools.
This really resonates. I’ve had variable results with ChatGPT, Claude and Perplexity with each showing different strengths and weaknesses. Your tips are very useful indeed - thank you!
Would love to hear what strengths and weaknesses you've noticed -- especially if they're different from what I've described
This is the most comprehensive guide I have seen so far.
It’s definitely getting on my weekend reading list :)
Thank you! I'd be honored if someone read this on their weekend. I mean I wrote this thing, and even I would rather read a murder mystery
Phenomenal stuff! I've only dabbled in deep research a little bit and got mixed results, I'm definitely going to take your learnings and try to squeeze out more useful things out of chatgpt. Thank you!
In my experience Deep Research can really range from "meh" to mind-blowing, depending on 1) the topic, 2) the tool, and 3) the context / prompt you provide. Definitely worth playing with more
Good thorough post. Very useful 👍🏼
Good article. How do you structure your prompts to get great results
Amazing resource, thank you Torsten!!
I read this, then I double check if I'm subscribed (even though I already subscribed a while ago).
If this is the quality you write regularly, I really don't want to miss them.
Haha, thank you, that means a lot! I try to make every post as useful as this one. Stay tuned for the next ones :)
After the first round I pass the finished research papers around to the other LLMs and have them critique them and improve them. I start the prompt with “You’re reviewer 2.” I find iterating tends to improve things, particularly if I noticed an error or weaker section somewhere.
Thanks for sharing - is it fair to say the investment required to train the models mean companies will carve out niches rather than resulting in one model to rule them all?
For academic purposes I purchase Consensus, which solves most of the bias, citation, accuracy issues.
Such a well written and helpful piece. Thank you!!
Might be good to specify how you want the citations formatted, since the output might otherwise display them too tersely.