Plain English is not the only notation you can hand a model. Between free-form prose and raw code sits a family of constructed notations we call PseudoLangs — purpose-built ways of talking to AI. We introduced them here as a family; the intervening research has given us solid ground to say which ones earn their place, where, and why.

The starting premise turns out to be well-founded: how you encode a prompt — not just what it says — materially changes the result. Research on prompt-format sensitivity found that meaning-preserving changes to formatting alone can swing a model's accuracy by a large margin. PseudoLangs are the deliberate use of that lever.

A spectrum, not a grab-bag

The family spans a spectrum by how far it travels from natural language:

  • PseudoScript / SudoCode — structured pseudo-code: numbered steps, named functions, variables, explicit control flow. Moves toward precision by making procedure explicit.
  • MiniScript — compressed, minified prompting that drops filler to fit a token budget. Moves toward efficiency.
  • SymboScript — symbolic encoding that packs meaning into glyph sequences. Moves toward density.

Read it as a dial. Turn it toward precision and you get pseudo-code; toward efficiency, compression; toward density, symbols. The first is well-evidenced and production-ready, the second is real but best automated, and the third is an experimental frontier with hard limits. Knowing which way to turn the dial — and how far — is the skill.

What the evidence says about each

Structured prompting is the strongest member. Prompting with Pseudo-Code Instructions (Mishra et al.) rewrote 132 tasks as pseudo-code and found it beat natural-language prompts by a wide margin, because pseudo-code reduces the ambiguity prose carries by default. A complementary study, Code Prompting Elicits Conditional Reasoning, traced the gain to improved variable-state tracking. This is PseudoScript, validated.

Compression is real but should be measured. Tools like LLMLingua show that dropping low-information tokens can compress a prompt dramatically for a small, measurable accuracy cost — the production-grade version of MiniScript's hand-minification.

Symbolic encoding is the frontier. The intuition that symbols are a free compression is mostly wrong: most emojis tokenize into multiple sub-tokens and models read uncommon symbols unreliably. The principled version is to let the model learn compact tokens, as in gist-token compression — not to bolt glyphs onto inference.

The line that keeps it honest

A structured notation is legitimate when the model acts on it and a human can maintain it. It becomes theatre when it is decorative — syntax that describes a process the model never actually executes, layered on to look rigorous. Keep the structure the model follows; cut the notation only you perform. The articles that follow take each member of the family in turn.