Do you have recommendations on specific non-negotiable aspects for prompts that product developers should keep in mind? You probably published a resource on this earlier, but I seem to have missed it.
Great question, Vishal! The best practices for prompts are evolving as the models advance. I believe there will be fewer non-negotiable aspects in the future.
Generally speaking, providing (necessary and well-defined) context and be relatively (clear and specific) about your ask are the most important aspects in my opinion.
Then there are nuances around which AI model you are using.
For reasoning models like o1, it's helpful to keep your prompt simple and concise. And be succinct with the context you provide to avoid it overthinking.
For other models like 4o and Sonnet 3.5, the basic techniques are still effective to improve the result: provide step-by-step instructions ("chain-of-thought prompting"), format your prompt properly, include examples, assign personas, include output template, etc.
This comparison is GOLD! Most articles hype one tool, but seeing all three tested with the exact same prompt reveals so much. Bolt surprisingly outperformed the others in functionality, which contradicts a lot of the online chatter favoring V0.
The speed vs quality tradeoff is fascinating - Lovable was lightning fast but basically non-functional! Really proves we need to test these tools ourselves rather than trust the hype.
Sure, Nick! Claude is quick and versatile in generating a mockup based on a prompt. However, the quality is usually not as great as app builders like V0, Bolt, or Lovable. Still, it’s helpful to use Claude to quickly check whether your prompt makes sense before using V0, Bolt, or Lovable, as your prompt might be unclear, too broad, or convoluted.
If you think the generated result from Claude is on the right direction regarding design/functionality, then you can go ahead to feed the prompt to those app builders. Using Claude as a proof-of-concept test can help prevent wasting tokens in those tools, eventually getting what you want more efficiently.
That’s what I suspected but I hadn’t realised Claude could produce these kind of mock-ups. I know it can write code but didn’t know it could visualise the code too
Working on the prompt outside these prototyping tools makes a tonne of sense
Great to discover your content from a Google search this morning 🙌
Do you think the Markdown prompt makes a difference? I tried “casual” prompting on Cursor - just provided a general idea and high level user flow - and it could generate relevant fields and even features without explicit instructions.
Btw aigents.pm (The Product Compass) has a PRD generator!
When you said Cursor, did you mean to say Claude? Or you did try Cursor out. That's impressive.
Thank you for sharing aigents.pm! I'll check it out. A few students of my courses mentioned https://www.chatprd.ai/, but I haven't got a chance to try that too.
In my case, I built a GPT specifically for clarifying ideas and generating prompts for AI prototyping tools rather than creating a PRD, as some parts of PRD are irrelevant.
Learned something new, I will try markdown sometime, thanks!
I meant Cursor. Pair that with Vercel and Firebase, very user-friendly. Deployment is swift.
Aigents.pm is for product managers and have a few nifty tools including generating assumptions (I have used ChatGPT for that and find that sufficient as well).
Love the idea of your custom ChatGPT for AI prototyping tools.
Interesting. Do you have any prompt hacks that can improve the UI because I feel like most of them look so basic and plain?
Have u taken similar cases with Replit and Heyboss?
Thanks for the message, Wyndo. An immediate tip I can think of is to provide it with references, such as snapshots of other UIs.
If you don't have any references, you can try tools like Galileo AI to generate wireframes and use them as a reference. I used UX Pilot for a similar workflow and shared in my last week's newsletter: https://designwithai.substack.com/p/how-to-turn-an-idea-into-a-prototype.
Tools like V0 started to integrate with local Design Systems—so I think it's just a matter of time to see the advancement in this area.
I have not taken similar cases with them. I've been busy recently—will take a look sometime. Please keep me posted if you get a chance to explore!
Nice breakdown, Xinran.
Do you have recommendations on specific non-negotiable aspects for prompts that product developers should keep in mind? You probably published a resource on this earlier, but I seem to have missed it.
Great question, Vishal! The best practices for prompts are evolving as the models advance. I believe there will be fewer non-negotiable aspects in the future.
Generally speaking, providing (necessary and well-defined) context and be relatively (clear and specific) about your ask are the most important aspects in my opinion.
Then there are nuances around which AI model you are using.
For reasoning models like o1, it's helpful to keep your prompt simple and concise. And be succinct with the context you provide to avoid it overthinking.
For other models like 4o and Sonnet 3.5, the basic techniques are still effective to improve the result: provide step-by-step instructions ("chain-of-thought prompting"), format your prompt properly, include examples, assign personas, include output template, etc.
Thanks. I learned something new today 😊
This comparison is GOLD! Most articles hype one tool, but seeing all three tested with the exact same prompt reveals so much. Bolt surprisingly outperformed the others in functionality, which contradicts a lot of the online chatter favoring V0.
The speed vs quality tradeoff is fascinating - Lovable was lightning fast but basically non-functional! Really proves we need to test these tools ourselves rather than trust the hype.
For anyone building quick prototypes, I explored similar challenges here: https://thoughts.jock.pl/p/building-apps-with-ai-2025-late-night-coding-adventure
Thank you for actually clicking buttons and testing functionality instead of just admiring pretty interfaces!
Appreciate it, Pawel! I look forward to your future explorations in this space too.
Thanks for this write-up. One question - can you explain further what you mean by using Claude as a POC test?
Thanks
Sure, Nick! Claude is quick and versatile in generating a mockup based on a prompt. However, the quality is usually not as great as app builders like V0, Bolt, or Lovable. Still, it’s helpful to use Claude to quickly check whether your prompt makes sense before using V0, Bolt, or Lovable, as your prompt might be unclear, too broad, or convoluted.
If you think the generated result from Claude is on the right direction regarding design/functionality, then you can go ahead to feed the prompt to those app builders. Using Claude as a proof-of-concept test can help prevent wasting tokens in those tools, eventually getting what you want more efficiently.
That’s what I suspected but I hadn’t realised Claude could produce these kind of mock-ups. I know it can write code but didn’t know it could visualise the code too
Working on the prompt outside these prototyping tools makes a tonne of sense
Great to discover your content from a Google search this morning 🙌
Oh yes, Claude can visualize the code and it does a good job.
Surprising to hear! Didn't expect that. Always thought Substack is not very SEO-optimized.
Thank you!
Do you think the Markdown prompt makes a difference? I tried “casual” prompting on Cursor - just provided a general idea and high level user flow - and it could generate relevant fields and even features without explicit instructions.
Btw aigents.pm (The Product Compass) has a PRD generator!
I think so, Evelyn, although I haven't compared markdown and "casual" prompt head-to-head.
Here's a related Reddit thread about the usage of markdown in prompting: https://www.reddit.com/r/PromptEngineering/comments/17aktzb/do_you_write_your_prompts_in_markdown/
When you said Cursor, did you mean to say Claude? Or you did try Cursor out. That's impressive.
Thank you for sharing aigents.pm! I'll check it out. A few students of my courses mentioned https://www.chatprd.ai/, but I haven't got a chance to try that too.
In my case, I built a GPT specifically for clarifying ideas and generating prompts for AI prototyping tools rather than creating a PRD, as some parts of PRD are irrelevant.
Learned something new, I will try markdown sometime, thanks!
I meant Cursor. Pair that with Vercel and Firebase, very user-friendly. Deployment is swift.
Aigents.pm is for product managers and have a few nifty tools including generating assumptions (I have used ChatGPT for that and find that sufficient as well).
Love the idea of your custom ChatGPT for AI prototyping tools.
That's amazing. I haven't touched Cursor. I'm inspired that you've already tried it.
That's an intriguing site. Appreciate you sharing.
Thanks. I've been busy recently—will share that custom GPT and how I built it in a newsletter.