Natural language is a wonderful interface and a leaky one. Ambiguity, implied context, and loose structure that humans paper over without noticing are exactly the things a model can misread. PseudoLangs are our answer to that gap: small, constructed notations purpose-built for talking to AI, used alongside plain language rather than replacing it. This is the case for them — and what the research now says about when they actually help.
Why a constructed notation helps at all
The starting claim is no longer just intuition. Research on prompt-format sensitivity found that meaning-preserving changes to how a prompt is formatted — not what it says — can swing a model's accuracy by a large margin. In other words, the encoding is load-bearing. A PseudoLang is a deliberate choice about that encoding: trade some of natural language's looseness for structure the model reads more reliably.
The family, and where each earns its place
PseudoLangs span a spectrum from natural prose toward more formal encodings, and the evidence is different for each:
- PseudoScript and SudoCode — structured pseudo-code (steps, functions, control flow). The strongest member: pseudo-code prompts measurably beat natural-language prompts by reducing ambiguity, with the gain traced to better variable-state tracking.
- MiniScript — compression to fit a token budget. Real, but best done with an evaluated, automated compressor (LLMLingua) rather than by hand, and never on the instructions or output contract.
- SymboScript — symbolic/emoji encoding. The experimental frontier: most symbols cost more tokens, not fewer, and reliability degrades on uncommon glyphs. Promising in narrow, measured settings; not a general-purpose scheme.
The full overview, with the spectrum laid out, is in the introduction to PseudoLangs.
The one rule that keeps it honest
A constructed notation is worth using when the model acts on it and a human can maintain it. It turns into theatre when it is decorative — syntax that describes a process the model never actually runs, layered on to look rigorous. Keep the structure the model follows; cut the notation only you perform. Used that way, PseudoLangs are a genuine bridge: a deliberate encoding that closes the distance between what you mean and what the model does with it.
Comments