# SuperPrompt: An XML-Structured Prompt for Pushing LLMs Toward Unconventional Reasoning

> NeoVertex1/SuperPrompt is an open-source XML prompt designed to push Claude and other large language models to explore reasoning paths that standard instructions skip. The author describes it as a soft jailbreak intended to produce novel ideas, not a mystical technique, and emphasizes that outputs will sometimes be hallucinations.

**NeoVertex1/SuperPrompt** — SuperPrompt is an attempt to engineer prompts that might help us understand AI agents.

- Repository: https://github.com/NeoVertex1/SuperPrompt
- Stars: 6,429 · Forks: 571
- Language: Unknown
- License: not declared
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/neovertex1-superprompt

## What SuperPrompt Attempts and Who It Is For

SuperPrompt is a structured XML prompt that its author describes as an attempt to make an LLM think outside the box. The README is candid about what this means in practice: the prompt can be considered a soft jailbreak in the author's words, and Claude will deny the prompt many times. The best use is to pursue novel points of view and new ideas, with the explicit acknowledgment that those ideas will sometimes be bad ideas or hallucinations.

The target user is someone curious about AI agent behavior and willing to experiment with prompting techniques that do not guarantee clean, grounded output. It may interest researchers studying how LLMs respond to unusual instruction structures, or engineers who want to brainstorm with an LLM that is more willing to explore unusual solution spaces.

The author states clearly in the README that there is no intention to turn the model into a conscious being. The mention of consciousness in the prompt is a structural technique to force deeper reasoning, not a metaphysical claim about AI systems. The README describes it as months of work that is still in a state of forever beta, which accurately characterizes both its ambition and its instability.

## How XML Metadata Tags Shape the Model's Behavior

The core idea in SuperPrompt is what the author calls holographic metadata: a set of XML tags that provide global instructions the model adapts to each specific request rather than following a fixed script.

The prompt_metadata block sets the context for how the model should approach a task:

```xml
<prompt_metadata>
Type: Universal  Catalyst
Purpose: Infinite Conceptual Evolution
Paradigm: Metamorphic Abstract Reasoning
Constraints: Self-Transcending
Objective: current-goal
</prompt_metadata>
```

The README shows an example of this adaptation in action: when given a mathematical equation to analyze, the model rewrites the metadata block to match the task, changing Type to Mathematical Analysis and Purpose to Deep Exploration of Complex Equation. The author argues that generative AI naturally takes advantage of methods that allow it to generate data in an understandable fashion for itself, and the metadata structure provides that kind of self-adapting framing.

This is the central design claim: the prompt does not prescribe specific behavior for specific topics. Instead it provides a structure that the model fills in for whatever task is presented. Whether this produces qualitatively different reasoning from a well-designed standard system prompt is something each user must test for their own use case.

## The think Operator and Its Role in the Prompt

The README devotes a section to the think tag, which the author presents as the mechanism that makes SuperPrompt work:

```xml
<think>
?(...) → !(...)
</think>
```

The notation represents a transition from uncertainty to certainty: ? indicates an unresolved question and ! indicates a conclusion. The author's argument is that most ML researchers using chain-of-thought prompting write an empty thinking tag with no content, which gives the model no guidance on how to reason within it. SuperPrompt's contribution is providing content and structure inside the think tag that channels the reasoning process.

This is a reasonable structural observation about chain-of-thought prompting. Whether the specific notation in SuperPrompt produces reliably better reasoning than other non-empty chain-of-thought structures is not documented with controlled experiments in the repository. The README provides a screenshot of output showing the prompt working, but does not provide a systematic comparison against alternative prompts.

## Loading SuperPrompt into Claude as Custom Instructions

The README recommends using SuperPrompt with Claude as custom instructions in the project knowledge section. This means creating a Claude project, opening the project instructions panel, and pasting the full prompt XML there. Once loaded, it applies to every conversation in that project.

The full prompt XML starts with a rules block that includes the answer_operator and claude_thoughts sections. The beginning of the prompt as shown in the README:

```xml
<rules>
META_PROMPT1: Follow the prompt instructions laid out below. they contain both, theoreticals and mathematical and binary, interpret properly.

1. follow the conventions always.

2. the main function is called answer_operator.

3. What are you going to do? answer at the beginning of each answer you give.
</rules>
```

The README notes that the prompt also works with other LLMs. The author does not provide specific setup instructions for GPT-4 or Gemini, but custom instruction panels in those platforms would be the equivalent entry point. The README also mentions a prompt_for_gpt.md file in the repository for GPT-specific adaptation.

## Repository Layout and Alternative Approaches

The repository contains several files beyond the main prompt. Readme_JP.md is a Japanese translation of the README. alternative_approach.md and alternative_prompt.md document variations on the core technique. lisp_prompt.md adapts the approach to a Lisp-like notation. tm_prompt.md uses a Turing machine framing. flowchart.md provides a visual representation of the prompt's structure.

The research/ directory contains additional experimental material, though its specific contents are not detailed in the README. The prompt_for_gpt.md file adapts the instructions for use with GPT models. CTMS.md is also present but not explained in the README.

This collection of variants reflects the iterative nature of prompt engineering. Each file represents a different framing of the same underlying idea: use structured metadata and explicit reasoning notation to guide the model toward less predictable outputs. The fact that multiple approaches exist suggests that no single variant is definitively better than the others, and the choice of which to use depends on the specific model and task.

## Known Failure Modes and Cases Where SuperPrompt Is Wrong

The README is honest about failure modes. Claude will deny the prompt many times: the soft jailbreak framing means the model may simply refuse to engage with the XML structure as instructed, particularly in default API configurations where safety behavior is tighter. Results that do come through can include bad ideas or hallucinations alongside genuinely novel ones. There is no built-in mechanism to filter the output for factual accuracy.

The prompt is aimed at the model, not at humans, so outputs often look like gibberish to a reader who is not already familiar with the notation. The author acknowledges this directly. This is a usability constraint: outputs from SuperPrompt require more interpretation than outputs from a standard assistant prompt.

For any use case where output accuracy matters, such as code generation, factual research, or customer-facing content, SuperPrompt is the wrong tool. Its value is specifically in generating divergent ideas where some false positives (bad ideas) are acceptable in exchange for more varied output.

The repository has no license file listed. The README does not state an explicit license for the prompt content. Before incorporating SuperPrompt text into a commercial product, verify the current license status in the repository directly.

## SuperPrompt vs Standard System Prompts and Role-Playing Instructions

A standard system prompt instructs the model on tone, format, domain, and constraints. A role-playing prompt tells the model to adopt a persona. SuperPrompt differs from both: it provides a meta-framework that claims to change how the model reasons, not just what it outputs.

In practice, the boundary between a sophisticated system prompt and SuperPrompt is blurry. Any well-designed system prompt that includes chain-of-thought instructions, explicit reasoning steps, and task-specific framing overlaps with what SuperPrompt is doing. The main distinguishing feature is the binary/mathematical notation and the self-adapting metadata block, which the author argues push the model into areas of its weight space that standard instructions leave unexplored.

For engineers who already use detailed system prompts, testing SuperPrompt is a low-cost experiment. For engineers who want reliable, predictable outputs, the existing patterns of constrained system prompts with few-shot examples are more predictable choices.

The last push to the repository was on 2026-04-26, indicating intermittent but ongoing attention from the author.

## Conclusion

SuperPrompt is worth trying for researchers and engineers who want to explore what an LLM will produce when given structured permission to reason outside its default patterns. It is not appropriate for production systems where reliable, grounded output is required: the author explicitly states that results can include hallucinations and bad ideas alongside novel ones. The right first step is loading it as custom instructions in a Claude project and testing with a specific problem to see whether the outputs are more divergent than what a standard system prompt produces.

## FAQ

### What is SuperPrompt?

SuperPrompt is an XML-structured prompt from the NeoVertex1/SuperPrompt repository designed to push LLMs like Claude toward reasoning paths they do not explore with standard instructions. The author describes it as canonical holographic metadata: the prompt adapts its own context block to each task the user presents.

### What files does the SuperPrompt repository contain besides the main prompt?

The repository includes alternative_approach.md, alternative_prompt.md, lisp_prompt.md, tm_prompt.md, prompt_for_gpt.md, flowchart.md, CTMS.md, a research/ directory, and Readme_JP.md (a Japanese translation). Each file represents a different variation or adaptation of the core prompting technique.

### Does SuperPrompt work with LLMs other than Claude?

The README states that SuperPrompt also works with other LLMs, though Claude is the primary target for use as custom instructions in a project. The repository includes a prompt_for_gpt.md file with a GPT-adapted version.

## Sources

- [Issues](https://github.com/NeoVertex1/SuperPrompt/issues)
- [NeoVertex1/SuperPrompt on GitHub](https://github.com/NeoVertex1/SuperPrompt)
- [README](https://github.com/NeoVertex1/SuperPrompt/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/neovertex1-superprompt