I gave Gemini Pro a pdf of the Wretched of the Earth and had it rewrite the story of the Grinch through that perspective.
But you don’t just do this in one prompt:
❌️
rewrite the story of How the grinch stole christmas from Fanon’s perspective (specifically The Wretched of the Earth).
❌️
You need to lead the LLM correctly to obtain good results, and for that you need to architect the process.
What I did was send Gemini (Pro, as it has 1m token context window) a simple PDF of the Wretched of the Earth, and then thoroughly prompted it:
✅
We are going to be rewriting the famous classic Dr. Seuss story “How The Grinch Stole Christmas”, but in this version, the Grinch finds a copy of the wretched of the earth - and everything changes.
Since this is a big undertaking, first you must isolate the main points from the Wretched of the Earth (provided here) and summarize them in a way that will fit in the story. You can keep track of all this internally, you don’t need to explicitly say it, but we can’t rely on the entire book because it’s a lot of text at once. So first, isolate what you think will be the most important lessons from the book for our Fanonized Grinch when we rewrite the story.
One thing I want to add: we will be picking heavily from Fanon and the copy I sent to rewrite our story. It will be an analysis of Fanon’s theory practiced in the story of the Grinch - so we will rewrite the Grinch as a short story (eventually when we reach that step), but taking heavily from Fanon to illustrate his concepts in practice/reality, right? It will be more of an analysis but in short story/prose/novel format.
Follow these steps in order to rewrite this story:
- First, go through the book provided and summarize/explicitize/write down the theoretical parts that will apply from Fanon to our new story. Also explicitize how they relate to the original story. There is a preface from Sartre in this book, ignore it. Only Fanon.
- Then, write a draft for each of the four acts (this story WILL be in four acts) that follows the Grinch’s story but from a Fanonian perspective. You are allowed to change the original story to fit this new framework.
- Finally, we will be writing out each act fully in a novel/prose style, with each part being about 250-300 words.
Let’s begin with step 1 when you are ready; take your time, be thorough, perform step 1 and let me know when you’re ready for step 2.
✅
The important part here is the 3 steps I laid out. You don’t actually even need to follow each step, it’s just that the LLM exists kind of in a vacuum of the context window so the ‘bigger’ you make your project seem, the more it has to draw on for this task.
Imagine you are giving this to a writer through email basically.
Deconstructing the prompt, we have the following structure:
- Clear outlining of the task and what is expected (we are going to do X)
- Scoping of the task: we know it’s a big job, so we will do it in a certain order.
- Additional instructions: I added these (the “one thing I want to add” part) because after a first few attempts I realized it was missing on some context. On web interfaces you do have this trial and error, on agentic software not as much (as the agent will handle the trial and error for you)
- The process to follow, with strict parameters. I relayed to the LLM several times that it must follow this order. This is more for Deepseek as it tends to want to do everything at once.
The reason I used Gemini Pro instead of my usual Deepseek is because Gemini has a 1 million tokens context window vs. deepseek’s 128k, and because the book is quite long (67k words), it was easier to do it
If you were to do it with smaller context windows, we can again break the task down in manageable steps:
- Open one conversation with deepseek, send first half of Wretched and same prompt - but explicitize it will only do acts 1 and 2.
- Open second conversation, send second half, same prompt, explicitize it will work on acts 3 and 4.
- Once you have the drafts with explicit Fanon references in it and where they apply, send the two to a third new conversation and have it work on step 3, the writing.
This is more akin to working with several people in a team. Everyone has their role and the deliverable is being built in a pipeline, passed off to the next person in the pipeline at each step to continue with their expertise.
In agentic, you are not so limited by context window - the software summarizes and resets a context accordingly. But I didn’t want to pay money for this so that’s why I used a web interface.
However, it still pays to steer agentic through these tasks to get a much better result. You become more of an architect that can break the problem down in manageable steps and then pass that on to the LLM. That initial part, setting the scope and architecture, is very important when working with LLMs. Maybe the next generation of LLMs won’t need it but for now, they still do.
In agentic the process would be much of the same: you give the agent a PDF of fanon in a folder, then give it more or less the same prompt but tell it to write down its final output for each step in a document (ANALYSIS.md, DRAFTS.md, FINALSTORY.md for example). That way when context window resets it can read those documents to catch up. It’s kind of like it has inherited a project from someone else and has its own documentation to understand the project instead of reinventing it from scratch.
-> You can go further with a critic. Still in agentic or web interface, still armed with the PDF of the wretched of the earth, you can do a new conversation/context, send it the draft, and ask it “You are Fanon, is this good, is this accurate, does it pass your critique?”. But that’s going much further than what I wanted for this fun retelling lol. You don’t have to do the critic but in some cases it does help improve the output.
And if you want the final story (I was more interested in the draft of the acts instead of a full story but I had Gemini make it anyway just in case): https://privatebin.luckvintage.com/?f2816bc405d90d7c#Ca15tpvN1iow8y36PrgUtPb8ctxVDEskPx7BA6SgQMJi (privatebin, like pastebin but not pastebin).
In conclusion, I hope this was followable for newcomers to agentic AI and stained glass Lenin wishes you a Merry Christmas comrades!