The 'forever beta' meta-prompt that speaks math to LLMs

What it does SuperPrompt is essentially one massive XML prompt intended as custom instructions for Claude (or other LLMs). It wraps your actual request inside layers of mathematical notation, binary strings, logical paradoxes, and pseudo-code loops that instruct the model to observe, analyze, and synthesize before answering. The author calls it a “soft jailbreak” meant to force deeper, less conventional reasoning, though he openly warns it can also produce bad ideas or hallucinations.

The interesting bit The prompt centers on a custom tischer tag and “holographic metadata” blocks that are supposed to dynamically adapt to whatever task you feed in—so asking about a complex equation supposedly shifts the model into “Mathematical Analysis” mode. The author argues this works because the prompt is optimized for the model’s own internal pattern-matching, not human readability, which is why it reads like a math department had a baby with an XML parser.

Key highlights

• Designed primarily for Claude, though the author claims it works with other LLMs.
• Uses a custom tischer tag and self-described “holographic metadata” to attempt task-specific adaptation.
• Explicitly labeled “forever beta” after months of iteration.
• The author openly admits Claude will often deny the prompt, and that outputs may be hallucinations or simply bad ideas.
• Not a consciousness hack: the author states the goal is deeper reasoning, not creating a sentient being.

Caveats

• “Forever beta” with no benchmarks or controlled evaluations; the metadata “adaptation” is shown only through a single screenshot and anecdote.
• Claude frequently refuses the prompt entirely, making it inconsistent to use.
• The author explicitly warns that novel outputs can be “bad ideas or hallucinations,” so the signal-to-noise ratio is unpredictable.

Verdict A fun rabbit hole if you treat prompt engineering as experimental art or want to probe the edges of Claude’s behavior. Avoid if you need reliable, production-grade outputs and don’t have time to filter noise.