While working with GenAI services (Perplexity, ChatPRD, ChatGPT, Gemini, Claude), feels very natural and conversational, it takes a little bit of work to get the right output. Here are the tips I’ve aggregated based on my usages and use cases:
If you are going to ask an LLM, it’s important that you choose a product that you can reasonably expect a good output. Here are my notes based on my experience so far:
Claude - Good for following directions in the prompt (add “”, tell me what you don’t know, etc). Also good for document-based RAG (retrieval augmented generation), due to the ability to upload documents to a session.
Gemini - Good for compliments! Use for the good news on critiquing activities.
Perplexity & Chat GPT - Good for starting. Generally better all around LLMs that can handle a wide variety of exercises without rate limiting.
Whether it’s just using text-to-speech, typing it out, or uploading a document, the more context, the better the output. Even past examples to scrape the format or capture the tone of the content, are extremely helpful.
Add a job title (expert Product Leader, CTO, etc) and voice instructions (make sure the tone balances being both professional and light). Even indicate who your audience is.
Variety is the spice of life. Especially if the content is short and the format is not well dictated, ask for multiple versions to choose from.
Especially when loading the prompt with information, adding “Don’t do anything with this information until I ask a question” will help focus the system while you load information and context. This can also be helpful when asking the LLM to break down a problem step by step. And, even when the GPT says they can support multiple tasks at once, better results have been reported by doing one task at a time.
Ask for critique on quality, tone and readability for things, including those that are outputs of an LLM.
There will need to be times that the output needs to be put in your own words. It’s important to know the difference of when this is necessary. As a rule right now, LLM creation should be made for DRAFTS, not final products with no human intervention.