Ess 0.2 · pseudo-code for LLM interpreters
Write it like notes. Every line carries a mark for what must hold, what's only a hint, and what's still open. A model can then turn it into real code without guessing, and tell you what it decided.
Ordinary pseudo-code never says how firm it is. "Sort the users" could be a strict rule or a placeholder, so an interpreter guesses, and guesses drift. Ess puts the firmness on the page.
Satisfy exactly. Turn it into a check or a test. Never trade it for anything.
Do what the line says. The interpreter picks the mechanism.
A hint. It can be improved on, and any change in behavior gets reported.
Not decided yet. The interpreter chooses, implements it, and says what it chose.
Context for people. Never a requirement.
When two lines disagree, the stronger one wins and the conflict is reported: ! › eg › step › ~ › style.
A mark governs the text after it, up to the next mark or #. Whatever comes before the first mark is a plain step. The same rule holds in terms, fields and parameters.
Examples are tests. Every eg, test and invariant line becomes an executable check in the generated code.
Terms fix vocabulary. term Slug = … means every later "Slug" is exactly that, so the rest of the file can stay short.
Decisions are reported. The interpreter returns code, tests, and a report of its decisions, deviations, conflicts and coverage.
Real output from the repository. wc and shortener were lowered blind, by a model that saw only the interpreter prompt and the .ess file. Point at a line on either side to see where it went.
Point at a line to trace it.
One line can stand for a whole function. Twenty lines can pin down every step. It's the same language either way: more detail just means less freedom. Write at the highest level that still pins down what matters.
To check that Ess can be interpreted from its prompt alone, three separate runs each got the prompt and one input, and nothing else. Their guesses became version 0.2.
| Run | Output | Tests | Result |
|---|---|---|---|
Lower wc.ess | Go command-line tool | 11 pass | Same output as GNU wc on files, totals, flags, a missing file and input with no final newline. |
Lower shortener.ess | TypeScript + Hono service | 22 pass | Includes 50 concurrent hits. Create, redirect, stats, 404 and 400 checked by hand with curl. |
Lift slugify.py | Ess, from code | – | Recovered every example and both invariants, and flagged two real bugs as ?. |
The blind lift marked as accidental that "a—b" became "ab", and that max_len was never checked.
Both became ! lines and eg lines in slugify.ess. Nearly 40 guesses across all three runs reshaped the spec.
The reference was regenerated against the tighter source. All tests pass, now including "a—b" → "a-b".
When lifting the lowered code gives back the same Ess, you've reached a fixed point: the source is down to its essentials.
Give any LLM INTERPRETER.md as its instructions, or use the bundled command-line tool, which calls Claude and writes the files out.
Read the report first. Every decision you want to keep becomes a ! or an eg; then lower again.
field?: type = defaultfail makes it fallible