This is an LLM prompt originally based on SuperPrompt for Claude, adapted to Gemini, and now forked, modified & extended. Additional features have been added to it and it's most powerful when used as a 2-stage pipeline. Feed it your problems and watch things unroll and expand! There are plenty example test cases available here. See here for very latest R&D.
Primary prompts:
- GEMINI-1.5_PRO.txt: The primary engine for problem exploration à la SuperPrompt. Use as 1st-stage.
Under active investigation:
- GEMINI-1.5_PRO.extra_gttm.txt: Additional expansion of quality of CoT thinking traces.
Less developed prompts, now parked:
- GEMINI-1.5_PRO.execution_engine.txt: A specialized generic secondary engine that processes the output from the primary engine to generate further insights and analysis. Use with the XML output of the primary engine. Use as 2nd-stage.
- GEMINI-1.5_PRO.schizoanalysis.txt: A specialized domain-specific engine designed for performing schizoanalysis on specific texts or concepts. This serves a niche use-case, so your mileage with it may vary. Use with the XML output of the primary engine or with your own texts / problem statements. Use as 1st-stage or as 2nd-stage.
- GEMINI-1.5_PRO.tm.txt: A first-stage primary engine, port of the Transcendent Mind Protocol from upstream. Use as 1st-stage.
- GEMINI-1.5_PRO.unrolled.txt: Like GEMINI-1.5_PRO.txt but generates additional local context. Use as 1st-stage.
With Google AI Studio or with the Gemini API. Make sure to put the prompts into the "System Instructions" box. All work here has only been done with the Gemini 1.5 Pro Experimental 0827 (2024) model. Using any other model, your results may vary.
Important: Gemini 1.5 Pro 002 is rather lazy. I strongly recommend that you use the experimental models instead. Here are your options:
(A): Gemini 1.5 Pro Experimental 0827 (2024) is no longer available in Google AI Studio. It is still available via API. Top tip: Vertex AI has better rate limits than Gemini on Google AI does for 0827. Consider using as a front-end: https://github.com/umaplehurst/gemini.pytk
(B): If using Gemini Experimental 1121 (2024) you will need to issue multiple "Please continue: [insert last line of previous output]" prompts as the model lazily terminates before producing the full output; this is because its runtime has been clipped to approximately 60 seconds per request.
When working with the 1st-stage, you can often "stay" in the same chat without having to adjust any System Instructions or open a fresh context. For example, if you say to Gemini, after it has produced the XML for the problem analysis, "please produce a 1-pager with the most salient insights from the XML", it will often give a well-formatted (non-XML) reply in the same context.
When pasting the output from the 1st-stage into the 2nd-stage, wrap the XML output
from the 1st-stage in ```xml ... ``` markup, like this:
```xml
<answer_operator>
<game_plan type="descriptive">
...
</gemini_thoughts>
</answer_operator>
```
Notes on model settings:
- "JSON mode" and "Code execution" should both be set to OFF.
- You need to set the right temperature. I've mostly been working with 1.1 to 1.5 temperature. Depending on your problem statement, you may want to experiment and find what temperature works best for you, depending on the type of output that you want to see.
These engines are not one-shot instant solution machines; they are a catalyst for creative disruption. It doesn't provide final answers, it is not going to instantly solve hunger or create peace in the entire world, but it can help you break free from conventional thinking, discover unexpected connections, and generate novel ideas for addressing complex challenges. The prompts generate conceptual entropy — schizoanalytically, a surge of unpredictable ideas, connections, and lines of flight that can break you out of stagnant thought patterns.
By feeding your concepts, questions, or problems into the engine, you'll receive a blast of interpretations, unexpected juxtapositions, and thought-provoking challenges. This deliberate injection of entropy can destabilize your assumptions, spark new insights, and propel your thinking in entirely new directions.
These prompts are not just a tool, but are a process, for those who dare to think differently, to embrace the chaos, and to explore the uncharted territories of thought.
- (2024-09-07): "Beyond Linearity: A Schizoanalytic Approach to Emergence in Social Systems, Catalyzed by Large Language Models"
This project is an ongoing exploration, and we invite you to join the conversation.
- Explore the code and the documentation.
- Run the engines on your own texts or concepts.
- Saw some interesting results you want to share? Feel free to submit them as a test case.
- Share your feedback, insights, and criticisms.
- Contribute to the project through pull requests or issues.
A difference with this implementation is that the 1st stage is a lot more "verbose" about its output and tries to generate output for all the tags where it can. I find it to be far more useful this way, rather than in many cases keeping the tags "within" as SuperPrompt does for Claude.
Additionally:
-
Pay close attention to any "citation" hallucinations. This is where the LLM makes a "direct" citation to a URL (represented by a curly
”button in the AI Studio UI). See our paper above for further information on how this process works out in practice.If you get this, congratulations! This means that the LLM may have just synthesized a novel brand-new concept integration inside its latent-space representation, or, otherwise found an interesting direct link in its vector store.
My suggestion: do not dismiss it as "just" noise. Explore much further and deeper.