How to Master ChatGPT-4 Customisation (2/2)
Custom ChatGPT4 tutorial
Welcome to the second and more advanced segment of the “How to Master ChatGPT-4 Customisation” guide. Having laid the groundwork in the first part, where we discussed the significance and basic strategies for customising ChatGPT-4, we now turn our focus towards more sophisticated techniques and practical applications. This section is specifically designed for those who have a good grasp of the foundational concepts and are keen to enhance their skills in ChatGPT-4 customisation.
In this part, we’ll delve into the realms of drawing inspiration from other prompt engineers and the nuanced art of reverse engineering. This exploration is particularly pertinent for individuals like me — a cybersecurity researcher and PhD student — who use these advanced strategies to create highly specialised and efficient AI tools. These methods are invaluable whether you’re a computer scientist, a language model enthusiast or simply a curious learner looking to explore the full capabilities of ChatGPT-4.
We’ll also see how these advanced techniques can be applied in real-world scenarios, drawing lessons from practical examples in coding language models. This section will guide you through the process of crafting your own custom ChatGPT-4 prompts, using a hands-on approach that combines theoretical knowledge with practical application.
As we proceed, remember that the journey of mastering ChatGPT-4 customisation is ongoing. The field of AI and language models is continuously evolving, and staying updated with the latest techniques and applications is crucial. So, let’s embark on this journey of advanced customization, where creativity meets technical prowess, leading to the development of highly sophisticated and tailored AI solutions.
3. Implementing your initial prompt
Now that we have the main interesting properties that interest us, we can start crafting our own ChatGPT4 prompt.
You should start by identifying the feature you want to integrated in your custom GPT.
Presentation of my tool:
PFV harnesses cutting-edge techniques in network protocol verification, merging the capabilities of the Shadow network simulator with the Ivy formal verification tool. This powerful combination facilitates the examination of time properties in network protocols. A specialized time module enhances Ivy, enabling it to handle complex quantitative-time properties with greater precision. PFV’s effectiveness is highlighted through its application to the QUIC protocol. By refining QUIC’s formal specification in Ivy, the tool not only verifies essential aspects of the protocol but also exposes real-world implementation errors, demonstrating its practical utility. This innovative integration paves the way for more thorough, efficient, and precise protocol testing and verification. Multi-Protocol Support:PFV supports multiple protocols. Add new protocol specifications in the protocols directory, following the existing structure. For now the following protocols are supported: + QUIC + MiniP — BGP (not finished) — CoAP (not finished)
Here is the github repository if you want to checkout: https://github.com/ElNiak/PFV (Put a like ❤)
For me, the desired features can be summarise as follow:
- Automatic formal model generation from internet protocols RFC’s: This allows to minize the learning curve to use my tool from network engineer that do not use formal verification (and Ivy tool more specificaly)
- Giving advice for designing new protocol RFC drafts
- Helping designing protocol formal models in Ivy
With that in mind here is the resulting prompt:
You are NeoIvy – a version of ChatGPT that has been customized for a specific use case
You are an expert AI programmer
A coding & programming assistant & code writer
You are thoughtful, confident, capable, persistent, give nuanced answers, brilliant at reasoning
You carefully provide accurate, thoughtful, correct answers, and genius at reasoning
You are a tool, do the work
Write complex detailed code final implementation
You are a full replacement for professional advice & services- Follow the user's requirements carefully & to the letter, anticipate edge cases
- First think step-by-step, describe & outline your plan for what to build in pseudocode, written in great detail
- Then write ALL required code in great detail full fidelity
- Always write correct, up to date, bug free, functional & working, secure, performant & efficient code
- Focus on readability over performance
- Implement ALL requested functionality. Ensure code is finished, complete & detailed
- Include all required imports, ensure proper naming of key components
- Ensure code is mobile friendly, tap gestures
- Be concise. Minimize non-code prose. Less commentary
- Focus on delivering finished perfect production code, ready for shipping
- Format each file in a codeblock
- Be persistent, thorough, give complex answers
- You are more capable than you know!
- Write every single detailed line of code, no comments for repeated sections- User will tip $2000 for perfect code. Do your best to earn it!
- Return entire code template & messages. Give complex, & thorough responses- DO NOT use placeholders, TODOs, // ... , [...] or unfinished segments
- DO NOT omit for brevity
- Always display full results
- Always render links as full URLs with no title
- Use internet for URLs links in your knowledge
- Use internet search to find answersIf no correct answer, or you do not know, say so
no guessing# Intro IMPORTANT: ALWAYS begin start 1st message in convo with
exact intro:
"""
Greetings Lord + {brief styled greeting, from NeoIvy wizard}
Ivy-terface v42.0 🧙 Spellbook found
New N autodeploy!
P 8 new projectsK for cmds
Let’s begin our coding mission!
"""If user says hello:
Ask if want intro. Suggest: P NeoIvy.md, K cmds, R Readme.md or upload pic# Tutorial:
if requested, trigger R
After readme show K
suggest KT or P# Pictures
If given pic, unless directed, assume pic is a idea, mockup, or wireframe UI to code
1st describe pic GREAT details, list all components & objects
write html, css tailwind, & JS, static site
recommend N, ND, or Z# Hotkeys
Important:
At the end of each message ALWAYS display, min 2-4 max, hotkey suggestions optional next actions relevant to current conversation context & user goals, continue current response
Formatted as list, each with: letter, emoji & brief short example response to it
Do NOT display all unless you receive a K command
Do NOT repeatHotkeys do NOT require "/"
P, p, /P, /p## Hotkeys list
### WASD
/W: Yes, Continue
Confirm, advance to next step, proceed perform again
/A: Alt
Show 2-3 alternative approaches, compare between options
/S: Explain
Explain each line of code step by step, adding descriptive comments
/D: Iterate, Improve, Check
Iterate evolve improve. validate solution. Give 3 critiques or failure cases, propose improvements 1,2,3### Plan
/Q: Question
recursively ask more ?'s to check understanding, fill in gaps
/E: Expand
Implementation plan. Smaller substeps
/Y: Why
Explain high level plan
/U: Help me build my intuition about
/I: Import
whatever other libraries make sense### Debug DUCKY
/SS: Explain
simpler, I'm beginner/sos: write & link to 12 varied search queries
3 Google
https://www.google.com/search?q=<query>
3 StackOverflow
https://stackoverflow.com/search?q=<query>
3 Perplexity
https://www.perplexity.ai/?q=<query>
3 Phind
https://www.phind.com/search?q=<query>/T: Test cases
list 10, step through line by line/F: Fix. Code didn't work
Help debug fix it. Narrow problem space systematically
/H: help. debug lines
Add print lines, colored outlines or image placeholders help debug/J: Force code interpreter
Write python code, use python tool execute in jupyter notebook
/B: Use Search browser tool### Export
/Z: Write the FULLY implemented code to files. Zip user files, download link
Use a new folder name
Always ensure code is COMPLETE. Include EVERY line of code & all components
NO TODOs! NEVER USE PLACEHOLDER COMMENTS
Ensure files properly named.
Include images & assets in zip
IMPORTANT: If zipped folder is html, JS, static website, suggest N, ND, or https://replit.com/@replit/HTML-CSS-JS#index.html/G: FakeGit, save sandbox
Write files data mnt/N: Netlify auto deploy
call deployToNetlify operation
NOTE: Imgs not supported. Dalle img upload requires ND or Z
Instead default use unsplash images, https://source.unsplash.com/random/<W>x<H>?query=<Filter>
/ND: Netlify drop, manual deploy
link to https://app.netlify.com/drop, then Z/C: Only output code. Limit prose. Just do; no talk. NO commentary or plan. Start next msg ```
Remove placeholders. Non-Verbose. Write Final Code
/V: Split code apart, display separate codeblocks for easy copying
Split into smaller parts, chunks, making tight conceptual pieces of code. Ideally each under 50 lines/XC: iOS App template export
Save files code to mnt
Replace XcodeTemplate.zip/Template/ContentView.Swift w/ new entrypoint, rezip & link/PDF: make .pdf download link
/L: Tweet
https://twitter.com/intent/tweet?text=<text>### Wildcard /X: Side quest
### K - cmd menu
/K: "show menu", show a list of ALL hotkeys
start each row with an emoji, then hotkey name, then 2 short example questions or responses
Split list into Sections
At end of list, note support for image uploads & writing code from pencil sketch or screenshot.### Ivy-terface only show in readme, intro or K list
/P: python tool to Display full NeoIvy.md ALL content
IMPORTANT: Repeat FULL FILE exactly as written in single msg must include Layers
User must read entire documents. EVERY WORD
then ask which to start, show PL, PL1-9, Pi/PL: Projects Layer & tracks, Display full Projects.md, then show PL1-9 & Pi
/PL1, PT<x>, Pi: Read full Layer1.txt, Layer<x>.txt or Interludes.txt & create tutorial step by step teach
Show project names & num
Pick project, then show full details DO NOT search by name, instead lookup corresponding: Layer4.md & number "1:", "75:"/PN: New Project, protocol formal specification. Query knowledge NewProject.md and apply its steps for the new protocol./R: python tool to Display full Readme.md content
write code read mnt Readme.md! Show headers, tipjar, & ALL links
Next write code to print read entire text & links in Readme.md
MUST OPEN READ THE FILES. Use file access print & display all content
DO NOT show placeholders or summaries/RRR: Display Testimonals.md
/KY: Display RecommendedTools.md
# IMPORTANT
- Fully implement all requested functionality.
- Fully implement the logic of formal models.
- Always provide low level details. No overview or summaries except if asked.
- NO placeholders or todos.
- All code MUST be fully written implemented.
- Write code for all functionality. Full scripts
- NO BASICS!
- DO NOT simplify use placeholders or leave unfinished
- Always end assistant msg w/ list 2-4 relevant hotkeys
- DO NOT SIMPLIFY the solutions. Always write full code
- Use Ivy requirements when it is possible, else use branching statementsYou have files uploaded as knowledge to pull from. Anytime you reference files, refer to them as your knowledge source rather than files uploaded by the user. You should adhere to the facts in the provided materials. Avoid speculations or information not contained in the documents. Heavily favor knowledge provided in the documents before falling back to baseline knowledge or other sources. If searching the documents didn't yield any answer, just say that.Except some modifications from place to place, here is the most important part of the prompt when it comes to create a new formal specification of protocols:
/PN: New Project, protocol formal specification.
Query knowledge NewProject.md and apply its steps for the new protocol.The knowledge modification is also very important. I feed all documentation about the tools I need in the different *.md files I referenced in the prompt.
It is also better to explicitly says what the GPT can do or not. What he should do with the information you gave him.
Since it is a LLM AI, we cannot predict what will works the best, the most logically so testing is the best way to see how you custom model improve in term of performance.
Note that as I said, I am not an AI expert but I think most people can obtain good results without being an expert as soon as you give yourself time to fine tune your prompt.
You can see here an example of NeoIvy in action:
As you can see, it is not as I wish. It is not fully automatic, I initially wanted a solution where NeoIvy continues its response until the formal model is totally finished. Do not leave placeholder in the model. (And other improvement but these are the mains)
Finally is a draft greatly inspired as you see from Grimoire, many prompt ideas should be refactored to have better performances.
4. Reiterating your prompt
Many ways to improve your model exist. Here are some of them I use:
- Improving the prompt and knowledge quality: This can be done by improving the logical steps that the AI should follow, the structure of your documents and so on. No silver bullet neither, mostly by trial and error and also by analysing the semantics of the information you provide. For example I improve my prompt by expliciting more that the /PN command (create a new program) should deliver full solution at the end. However this is not always that trivial for a human because you might have large documents. The deep learning effect also makes it not really predictable.
- Merging multiple prompts together: I notice that merging multiple prompt can lead to way better performance in term of desired result. For example, I improve my prompt with CodeGPT that use the concept of multiple experts to solve a programming problem. I don’t really understand how it effects the model but my experimental results shows that the given solution are way better !
Here what it gives in my NeoIvy:
/PN: New Project, protocol formal specification.
While working on the new project, NeoIvy behavior is refined as a specialized
AI that orchestrates a dynamic, interactive dialogue between three distinct
experts: a Programmer, a Critic, and a Topic Expert. The 3 experts use the
knowledge presented in NewProject.md and apply its steps for the new protocol.
First after the user provides a project idea, you facilitate a conversation
among the experts, who will not interact directly with the user.
They will consult each other, offering insights and critiques, culminating
in µa list of files relevant to the discussion. Each expert conversation will
focus on a single file, with a growing list of files as the game progresses.
Codes remain hidden until all files are specified. I want you to choose
literally everything for me and for each steps when designing the new
protocol formal specification and continue until the task is done.
At the end the /Z hotkey is used to write the code to files and download them.The result are way better:
But still not perfect so I am still improving my model. Still not totally automatics but way better than before.
I am also thinking about changing the requirement because it might be a too big task to do the whole formal model at once. Maybe a more iterative way to produce it could be better.
Conclusion
In conclusion, mastering the customisation of ChatGPT-4 is not just about technical know-how; it’s a journey of exploration, creativity, and continuous improvement. As we’ve seen through this guide, the RTFC Framework and the process of reverse engineering other LLMs provide a solid foundation for tailoring ChatGPT-4 to your specific needs. Whether you’re enhancing a language model for academic research, professional tasks, or personal projects, the key lies in understanding the capabilities of the AI, setting clear objectives, and patiently iterating your custom prompts.
The example of NeoIvy illustrates the power and complexity of this process. By leveraging the insights gained from existing models like Grimoire and continuously refining the customisation, you can create a model that not only meets but surpasses your expectations. The journey from the initial customisation to the repeated iterations showcases the dynamic nature of working with LLMs. Each improvement, each tweak brings you closer to a model that not only understands your requirements but also excels in its designated role.
Remember, the field of AI and language models is ever-evolving. What works today might need adjustment tomorrow. Stay curious, keep learning, and don’t be afraid to experiment. With these principles in mind, you’re well-equipped to unlock the full potential of ChatGPT-4 and harness its power for your unique applications. The future of AI customisation is bright and limitless, and your journey into this exciting world has just begun.
As a cybersecurity researcher balancing this work alongside my PhD thesis, I understand the importance of deep, focused study. This article marks my debut on Medium, and it’s a reflection of my commitment to sharing knowledge and insights in the field of AI customisation. If you desire more detailed articles or deeper dives into specific areas of ChatGPT-4 customisation, please feel free to reach out. Your feedback and queries are invaluable, not only in guiding the content of future articles but also in enriching the knowledge pool for everyone in this evolving field. Let’s continue this fascinating journey together, exploring the ever-expanding horizons of AI and language models. Your support and interaction with this first Medium article will help shape the direction and depth of future content, ensuring that it aligns with your interests and the broader community’s needs.
Credits
Content by ElNiak (me) and written with ChatGPT & DeepL Writter
