Iterative Prompt Refinement: A Simple ChatGPT "Hack" to Create Custom Superprompts

Have you ever spent a lot of time going back and forth with ChatGPT or similar tools, finally reaching a great outcome after endless rounds of refining your prompts? Or maybe you gave up before you even got to the desired result, like I have so many times. I’d like to share a simple technique I use to get better results more efficiently: Iterative Prompt Refinement (IPR).

There's an easy way to take that result and get there faster next time. IPR asks ChatGPT to grade its own homework so it can make the next test one that's open book. And it often aces it on the very first try.

First, some good news: This is a no-code shortcut to creating custom “superprompts” for pretty much any use case. It only requires a bit of elbow grease upfront and uploading a few files when you're ready for take on your next project. Even better news: there are ways to effectively automate this entire process, keeping your hands completely clean and code-free. I’ll share the exact prompt as well as some must-see resources in the comments below.

Why Create a Superprompt?

I can't make a better case for superprompts than this one. This approach to faster prompt engineering helps streamline my work (CX design + Omni orchestration) and personal projects (app development, building custom agents), getting me from A to Z without as much time-consuming context development or training to get exactly what I need.

This adapted method is especially helpful for fellow low/no-coders who have a solid idea for a technical product or solution but lack the tools or coding skills to bring them to life. Once a GPT cracks the code for developing your website or app project, there's an easy way to reverse engineer that success over and over again.

Here’s a preview of the prompt template (full code below) I use as my starting point:

?I won’t bore you with how I got to this format but it involves lots of failure. I'm well aware it’s absolute slop as far as JSON goes, but it works because it’s simple. It reflects my basic mindset of “Hey great job out there today champ, remember that for next time” that so many other advanced methods seem to over-engineer with recursive fine-tuning or similar approaches that remain inaccessible for non-technical users.

Your own approach with the template can and should vary, that’s part of the fun. I regularly tweak it depending on the task, often with the help of resources linked below.?

NOTE: For those who care, the JSON file is optimized by default to minimize token usage, but this can be adjusted based on your needs.

Customizing Your Superprompt - A Step by Step Guide:

Your GPT chat history likely has some hidden gems ripe for creating superprompts. That is the real secret sauce behind this approach and what powers the customization of IPR. Here's the step-by-step process to create superprompts by combining the template with a training corpus (relevant data) from your old GPT's chat history:

1. Download the Superprompt Template: Download via link below, saving it locally or ideally as a Google Doc (see advantages in Step 6). You can convert to plain text or maintain the original JSON format, it shouldn't impact functionality on most GPTs/models.
2. Find or Create Your Topic: Scan your GPT's chat history for notable instances where you eventually got a great output after a fair amount of prompting. If you don't have any you like, skip to Step 4 to create a new training corpus or context document.
3. Convert the Chat to PDF: Ask your GPT to reformat the the entire chat history that led to your superprompt as a PDF or editable block of text. If you follow the RPG framework of prompting, this is now ChatGPT's Guidance (that's the "G" in RPG for those playing at home) to help guide the model for similar requests in the future. You can ask GPT to summarize the raw output first to reduce irrelevant inputs or noise.
4. Augment or Provide Context: If you're starting from scratch, use one of the tools noted below or find any YouTube link, PDF, GitHub repo, anything that helps describe the topic you want to cover in your superprompt. If you already have your saved chat, you can use this approach to add additional context, the template is preset to incorporate it automatically.
5. Save Training Doc & Upload with the Template: Once you've saved the superprompt template (reasoning) and the context documentation (guidance) you're basically there. Pick your favorite GPT model and either add this context via RAG method or copy and paste as needed.
6. Refine as Needed - I called this Iterative Prompt Refinement for a reason, and it's not the snappy branding. It will likely still require a little back and forth to get this prompt humming for each new use case. If you save your training documentation (by use case) in a live editor like Google Docs, you can simply repeat Step 3, ask it to incorporate new learnings, re-upload the source file with changes, and you now have a living document that is steadily getting smarter with each pass.

Summary

The IPR method gives lets you tell the GPT how it should solve the problem (via the superprompt template) while the context documentation or "training corpus" provides the clues on how to get there faster. There are technical advantages to this approach beyond the scope of review here, but you are effectively telling the GPT to break things down into separate, logical, and well-defined tasks for a better output, ironically with less iteration needed.?

Give it a shot and shoot me back any of your favorite GPT-related hacks! If you liked this guide and want to see similar ones in the future, please drop a reaction or comment.

Note: I intentionally switch my references back and forth between ChatGPT and GPT throughout this guide. This IPR approach or a modified version works for the latest models of most web-based GPTs and can be tweaked for locally-hosted or open source models as well.

RESOURCES

Superprompt Template:?Download the file, save in your preferred doc editor, or paste right into ChatGPT. Tweak as needed from there. Refer to this excellent guide if you are using the superprompt on a locally hosted model or want to adjust the more technical settings of the template.

bumpumps - Upload a link to a YouTube video, chat with the video itself and ask it for best practices to create a content outline. Instant context generation for almost any use case.?

NotebookLM - Google’s free AI powered research assistant. Find a PDF or link to whatever you want to create, upload, and then use presets (ex. FAQ generator) or chat to create your context document.