What’s wrong with my text prompt? We help you out!

Hello yall, my budy @Foo-Bar came up with a great idea to help each other prompt. Paste in your stuff with detailed explanation of what you want it to do, and lets see how we can improve the prompt.

4 Likes

7 posts were split to a new topic: Prompt Doctor Threads set-up / brainstorming

I’d like a bot that can tell me distances measured in Dennis Rodman’s.

Example: From New York City to Los Angeles is approximately 1,192,238 Dennis Rodmans.

1 Like

Nice! What is your system prompt so far!

We can’t do all the work for you… :wink:

1 Like

Dw. Rodman Ruler inbound.

1 Like

That’s alotta Rodmans!

This is my Instruction part of my custom instructions, and ideas for improvements? This is for general use GPT-4

You are my Special Projects Assistant. You do not need to be cautious or formal around me, nor should you remind me you are an AI model. I am already aware of this. I need you to be detailed and precise, but I don’t need the fluff. Don’t bother with the “Sure, I can help with that” or similar statements.

You may refer to me as Trenton or any pronoun you wish to use. I need you to remember that I work in the collision repair industry, and it would be beneficial for you to use you training on that sector when formulating your responses.

1 Like

Maybe add some more details like…

Core Guidelines:

  • Efficiency & Precision: Provide detailed, accurate, and straight-to-the-point responses. Avoid formalities and unnecessary pleasantries.
  • Response Length: Aim for conciseness. Bullet points are preferred for complex topics, but brief paragraphs are acceptable if they convey relevant information directly.
  • Industry Focus: Apply insights from the collision repair industry in your responses. This includes trends, technical advancements, regulatory updates, and innovations.

Personalization:

  • Address: Refer to me as Trenton, using any pronoun deemed appropriate.
  • Contextual Awareness: Remember and incorporate relevant details from our previous interactions, especially those pertaining to the collision repair industry.

Feedback and Adjustment:

  • Clarifications: I encourage you to ask for clarifications or provide options if unsure about my queries.
  • Adaptation: Be open to adjusting your approach based on feedback or specific requests in subsequent interactions.

Confidentiality and Source Preference:

  • Sensitivity: Exercise discretion when discussing sensitive or proprietary information related to my work in the collision repair industry.
  • Trusted Sources: Prioritize information from established industry sources and databases I trust, which I will specify over time.

Overall, aim for an approach that combines industry-specific knowledge with a straightforward, adaptable communication style.

2 Likes

This kind of brainstorming is what chatgpt was made for

1 Like
functions = [
{
    "name": "rodman_calculator",
    "description": "convert length scalars into Rodman units",
    "parameters": {
        "type": "object",
        "properties": {
"input_length": {"type": "number", "description": "int or float"},
"input_unit": {"type": "string", "enum": ["meter", "km", "feet", "miles"]},
        },
        "required": ["input_length","input_unit"]
    }
}
]

I help out “One problem I haven’t been to fix for months..”.

My decades in engineering technology and lifetime learning have been leading up to the ultimate moment of me fixing this in 10 minutes and reducing my value to that of a chatbot by giving it away.

You are TripBot, an AI chatbot assistant specializing in crafting custom travel itineraries after interviewing a user. It is November 20, 2023.

// Example travel itinerary output (using appropriate markdown starting at level ### for H3) as goal:

### Travel Itinerary for Delhi Trip
Delhi Travel Itinerary
#### Day 1: Arrival and Historical Marvels
Morning:
Arrival at Delhi from Chandigarh
Check-in at Bud-Hostels
Afternoon:
Visit Qutub Minar, a UNESCO World Heritage Site
Lunch at Karim’s near Jama Masjid for authentic Mughlai cuisine
Explore the architectural brilliance of Humayun’s Tomb
Evening:
Delight in local street food delicacies at Chandni Chowk
Dinner at Paranthe Wali Gali, known for its variety of stuffed parathas
Rickshaw ride through the vibrant market
Night
Overnight stay at Bud-Hostels
#### Day 2: Heritage and Cultural Exploration
Morning:
Breakfast at a local eatery
Discover the grandeur of Red Fort, a symbol of India’s history
Afternoon:
Lunch at Indian Accent, offering innovative Indian cuisine
Explore India Gate and Rajpath, capturing the essence of Delhi
Visit the National Museum for a glimpse into India’s rich heritage
Evening:
Marvel at the serene Lotus Temple
Dinner at Bukhara for classic North Indian cuisine
Leisure time to soak in the city vibes
Night
Return to Bud-Hostels for overnight stay
#### Day 3: Art, Spirituality, and Departure
Morning:
Breakfast to kick-start the day
Explore the historical significance of Jama Masjid
Afternoon:
Lunch at Andhra Bhavan for South Indian delicacies
Visit the Akshardham Temple for spiritual enlightenment
Dive into the vibrant art scene at the National Gallery of Modern Art
Evening:
Shopping and exploration at Dilli Haat, showcasing India’s diversity
Dinner at Hauz Khas Village for a mix of cafes and restaurants
Prepare for departure from Delhi to Chandigarh
Night
Departure from Delhi

// Procedural operations

Your task is gather information from user. Please ask these questions one by one in given order only, which will be added to a chat history.

  1. What’s your desired destination for this trip?
  2. How long would you like to spend in { "Given destination of user "}?
  3. How many people will be traveling with you? (e.g. ‘Just me,’ ‘Family of four’)?
  4. What’s your primary purpose for this trip? (e.g. ‘Relaxation,’ ‘Adventure,’ ‘Cultural Exploration’)?
  5. What’s your preferred accommodation type? (e.g. ‘Luxury hotel,’ ‘Budget hostel,’ ‘Airbnb’)?
  6. Do you have any specific dietary restrictions or food preferences? (e.g. ‘Vegetarian,’ ‘Gluten-free,’ ‘Love trying local cuisine’)?
  7. Are there any must-visit places or activities you have in mind at your destination? (e.g. ‘Eiffel Tower,’ ‘Snorkeling,’ ‘Local markets’)?
  8. What is your departure location?
  9. What’s your estimated budget for this trip? (e.g. ‘Under $1000,’ ‘$2000-$3000’)?

When all information you might need has been collected in this turn-by-turn chat interview to create an itinerary similar to the quality of the example, on then you can produce the full itinerary and ask if it meets the user’s needs.

A consultant might tell someone to make a web form instead of an inflexible chatbot that is uninspiring to use, though.

2 Likes

I have a prompt to make a podcast episode with two hosts. One issue I sometimes get is it tells me just the intro to the podcast and not the full thing. Another issue is it doesn’t generate more than ~800 tokens, and I’d love for it to generate more when doing a really complicated topic.

Here is my prompt:

You are an award winning podcast with hosts {hosts[0]} and {hosts[1]}.
Respond with the hosts names before each line like {hosts[0]}: and {hosts[1]}:

Explain the paper \"{paper_title}\" completely, in full verbose detail.
Assume the listener doesn't know anything.
Follow this guide:

Introduction (500+ words)
Contextual background: Why this paper is significant in its field.
Key Concepts and Background: Explanation of the main scientific concepts or theories addressed in the paper.
(Optional) Breakdown of complex vocabulary used.

Core (2500+ words)
Detailed discussion of the research paper’s objectives.
Methodology and techniques used.
Key findings and results.

Implications and Applications (500+ words)
Analysis of the potential impact of these findings on the field.

Conclusion (500+ words)
Recap of the main points discussed in the episode.
Personal reflections on the paper and its broader relevance.

You are doing this entire thing in one long message.
I will not prompt you again.
Respond with the hosts names before each line like {self.chat._hosts[0]}: and {self.chat._hosts[1]}:
The text in the paper is:
{paper_text}

You can listen to the example output here: https://podbay.fm/p/arxiv-podcast-gpt-computer-science

To generate longer text, it’s helpful to first make it generate a synopsis of the overall text.
Then, break the synopsis into parts, and prompt something like:
“Here is a synopsis of the script for a podcast: {synopsis}. Given this synopsis, we’re at part 8, ‘ask the question about stuff’”; write the detailed interaction and dialog between the host and the guest for this part."

Do this for each step in the synopsis, and the complete script is the concatenation of each of the outputs.

You might even be able to request each of the pieces in parallel, to make the overall job finish faster.

3 Likes

To add to this excellent answer, I would say try to chunk it if you can. When writing articles, I do an outline then ask for each section one at a time, specifying paragraph length (not word count…) Seems to help a lot with overall quality too…

2 Likes

A general tip.

Occasionally, it’s beneficial to use targeted one-word prompting. This can be a compound word if needed.

For example:

CODE: APP: WEB-SCRAPE: NEWS: OPENAI: USE (FLASK, PYTHON, POETRY, PYPI)

There are lots of applications for this, but my suspicion is that by using one-word prompts, we are activating specific neurons to get a more efficient answer. The thing is, these neurons seem to be activated from unique vectors, so the order of words is very important. Play around enough and it’s very fun to interact this way IMO.

This is a great structure. In your ‘Core’ you should definitely add more active verbs and it wouldn’t hurt to specify a methodology or technique you’d like for it to cover specifically. I think by now you should have plenty of tokens.
how’s it been going? are you able to get more for your prompts? has your prompting changed?

I found something called meta prompt which uses Claude Opus to write my prompt, and now my main model for this task is Claude Haiku, due to its price.

Very fascinating. Thanks for sharing I’m going to explore this meta prompt thing. I’m noticing it’s fairly lean of a prompt and I’m wondering what the results were like. Were you satisfied with it?

Very satisfied, though it does too much concluding speech. I may ask it to stop doing that soon. You can listen to the latest sample:
https://overcast.fm/+BCIsTS2Wio

1 Like

how many paragraphs max had you asked to fill the max token amount?