ChatGPT Prompt Engineering DIY Research: Master Prompt Crafting Today!

A practical workflow for inventing new prompt sequences for ChatGPT and other LLMs is built around mining research papers, turning them into reusable “frameworks,” and then stress-testing the results on a benchmark problem. The core idea is low-friction: find relevant academic work, summarize it with LLM plugins, synthesize a prompt strategy framework from multiple papers, and iterate by running the resulting prompt chain against real tasks.

The process starts with paper discovery, using ArcSave.org to search for topics like “prompting large language model.” The workflow then narrows to papers that look promising after skimming—examples mentioned include “strategic reasoning with language model,” “prompt based tuning,” “short answer grading using one shot prompting,” “code prompting,” “a neural symbolic method,” and “encrypted prompts.” Once a target paper is selected, its PDF link is pasted into ChatGPT using plugins such as “Ask Your PDF” and “Link Reader.” A first prompt requests an in-depth PDF summary, and a follow-up prompt asks for step-by-step instructions on how the framework works. Those summaries are saved into a text file so multiple papers can be processed and stored together.

Next comes synthesis: the saved summaries are fed back into ChatGPT to generate a new prompt-sequence framework. The resulting structure is described as designed to enhance strategic reasoning—guiding decision-making through elements like value assignment prompts, belief tracking prompts, chain-of-thought-style reasoning prompts, “racing” prompts, cascade prompts, and demonstration prompts. The creator typically integrates at most two papers, then adds a second research source—“Oola GPT empowering llms with human-like problem solving abilities”—to enrich the approach with additional prompting templates aimed at generating better questions, thinking templates, step thinking, and critical thinking.

After assembling the combined research insights, the workflow shifts from research to creation. A new ChatGPT run (using default GPT-4, without plugins) is prompted to generate fresh ideas that the user can research further, then the paper-derived material is pasted in. A final instruction asks for a step-by-step prompt chain that “super enhance[s]” logical problem solving, explicitly using chain-of-thought reasoning and other prompt-engineering techniques while avoiding direct replication of the papers.

The test phase uses a classic benchmark: measuring exactly 6 liters with a 12-liter jug and a 6-liter jug. The generated five-step prompt sequence is executed, and results are mixed. Early steps produce wrong or confused reasoning—at one point the model behaves as if it must manipulate both jugs despite the task not requiring that. Later steps still fail to produce the correct solution consistently. A regeneration attempt yields a more coherent response, with the model acknowledging earlier confusion and the premise that it should manipulate both jugs. The takeaway is not that the first synthesized prompt chain always works, but that the research-to-framework-to-benchmark loop can reveal what prompt structures help, what they miss, and where iteration is needed.

The transcript also detours into Nvidia’s gaming-focused AI stack—Nvidia Ace for games—describing how Nemo (character language models), Riva (speech-to-text and text-to-speech), and Omniverse Audio2Face (facial animation from audio) integrate with Unreal Engine 5 and Metahuman. That segment reinforces the broader theme: prompt and model orchestration techniques can be applied beyond ChatGPT to other AI systems and domains.