cd /news/ai-tools/natural-programming-just-dropped Β· home β€Ί topics β€Ί ai-tools β€Ί article
[ARTICLE Β· art-174353] src=github.com β†— pub= topic=ai-tools verified=true sentiment=↑ positive

Natural Programming Just Dropped

A new programming language called llang has been released, designed to be written by AI and read as structured, explicit English that behaves like Python. llang ships with an interpreter requiring only Python 3.10+, a WebAssembly backend for the integer/function subset verified by executing its output in Node, a self-hosted lexer written in llang and checked token-for-token against the bootstrap lexer, and a porting tool that converts Python, C and JavaScript repositories to llang with coverage reports and a JSONL cleanup queue. The language trades terse human-era syntax for labeled arguments, explicit block terminators and phrase-named builtins so reviewers can verify AI-written code, and includes native domain libraries for math, datascience (NumPy/pandas semantics) and machinelearning (scikit-learn semantics).

key takeaway

llang's explicit, labeled, terminator-delimited syntax is intended to make AI-generated code reviewable, and its porting tool converts Python, C and JavaScript repositories to llang with coverage reports and a JSONL cleanup queue for AI-assisted completion.

read6 min views7 publishedNov 22, 2026
Natural Programming Just Dropped
Image: Michielbdejong (auto-discovered)

A programming language that reads like structured English and behaves like Python.

llang is built to be written by AI. It is deliberately highly interpretable and direct: every construct announces itself, every block has an explicit terminator, every argument is labeled, and nothing hidden or implicit can slip past a reviewer. The goal is not clever code β€” it is code you can read once and understand, so humans can review what an AI wrote with confidence.

Human-era compromises (terse syntax, abbreviations, operator soup, clever implicit behavior) are pure downside for a model, so llang trades them for verbose, explicit, regular constructs. Natural-language phrasing is the point, not a novelty β€” llang is structured pseudocode you can actually run.

set scores to a map with "Ada" mapped to 91, "Grace" mapped to 98

repeat for each student in (call keys of with collection = scores)
    show student plus ": " plus (call text of with value = item student of scores)
end repeat

The language is implemented and usable. The primary execution path is the interpreter; a WebAssembly backend exists for the integer/function subset and is verified by executing its output in Node. For rewrites, "compiles" means the program parses, lints clean, and produces output identical to the original.

  • Values: numbers, text, booleans,nothing , lists, maps, functions
  • Statements:set /let ,show /display /print ,require ,add ,set item ,if /otherwise ,repeat while ,repeat for each ,stop the loop ,skip to the next item , functions with aliases, labeledcall ,fail /expect problem handling,use modules
  • Expressions: strictly left-to-right with parentheses for grouping, word operators with closed synonyms (floor divided by included), postfix predicates, collection literals, indexing, phrase-named builtins,apply for calling function values
  • AI ergonomics: definitions may use phrase names (define function find all ... ), and@ escapes reserved words anywhere a name is expected (taking @end ,call f with @end = 1 )
  • Domain libraries:advanced_math (math/statistics),datascience (NumPy/pandas semantics),machinelearning (scikit-learn semantics) β€” all native llang, with self-describing results and loud validation
  • Libraries: native llang modules inlib/ (math ,text ,lists ,shell ,shlex ,fnmatch ,textwrap ,csv ,json ,re ,ast_contract ,pylev ,humanize ), plus an explicitly labeled opt-in foreign bridge (json ,math ,shell )
  • Objects: minimal classes with explicitthis , single inheritance-free construction viacall new ... , andask instance to method ...
  • Compiler: a WebAssembly backend for the integer/function subset β€”llang-compile file.ll -o file.wasm , verified by executing the output in Node
  • Self-hosting: the language's own lexer is written in llang (lib/llang_lexer.ll ) and verified token-for-token against the bootstrap lexer
  • Universal porting app: point it at a repository and it converts supported source files (Python, C, JavaScript) to readable llang, preserving structure, with coverage reports, verification, and a JSONL cleanup queue for AI-assisted completion β€”llang-port-repo <repo> -o <out>
  • Tooling: source-located diagnostics, canonical formatter, pre-run linter, LSP language server, Python-to-llang converter, property-based differential fuzzer, single-command verification, GitHub Actions CI
  • Ecosystem: nativejson module, JSON package manifests with a registry, editor grammars for VS Code and tree-sitter, and a browser playground

The interpreter needs Python 3.10+ only. The full verification suite additionally uses Node.js (for WebAssembly checks) and the Python data libraries used as oracles: numpy, pandas, scikit-learn. Install them with pip install -e '.[dev]' before running tools/check.py.

python run.py examples/hello.ll          # run a program
python repl.py                           # interactive shell
python tools/check.py                    # every check: tests, lint, rewrites, links
python tools/port.py my.py -o ported.ll --report report.txt
python run.py --lint examples/hello.ll
python run.py --format examples/hello.ll
node tools/playground/server.js           # browser playground at localhost:8080
python -m llang.portapp.cli <repo> -o <out>  # port a whole repository to llang

Or install the package for console commands:

pip install -e .
llang examples/hello.ll
llang-port my.py -o ported.ll
llang-compile my.ll -o my.wasm
python -m llang.lsp            # language server over stdio

Surface flexibility, canonical precision. The surface language has a closed vocabulary of synonyms; the parser normalizes every spelling to one canonical AST node, and the interpreter only ever sees that node. Synonyms add readability without adding ambiguity β€” the failure mode that sank AppleScript and Inform 7.

Decision Rule
Semantics Grounded in authoritative docs of existing languages β€” re-express, never invent
Blocks Explicit end if /end repeat /end function /end expect ; indentation ignored
Evaluation Strictly left-to-right; parentheses for grouping
Types Dynamic; strict operations, no implicit coercion
Arguments Labeled roles ( call f with name = x ), never bare position
Extension Libraries extend values and types, never syntax
Diagnostics Source line/column and expected-token hints are part of the spec

Start here:

  • Tutorial β€” learn the language from zero
  • Reference β€” complete authoritative reference
  • Grammar β€” compact grammar that fits in an LLM prompt
  • Examples β€” runnable programs with pinned output

Design and process:

  • Design β€” decisions and rationale
  • Semantic decisions β€” tricky choices and why
  • Method β€” the repeatable process for adding a language surface
  • Family plan β€” tiered IR, converter, libraries, build order
  • Roadmap β€” honest scope and deferred features
  • Rewrites β€” open-source Python software rewritten and verified in llang
  • Research β€” grounded matrices: Python semantics, precedents, stdlib, demo, terminal, test plan

Project:

  • Changelog

  • Contributing

  • License β€” MIT

  • examples/report_card.ll β€” flagship program: a graded report card that checks its own result

  • examples/rubiks.ll β€” Rubik's cube simulator that writes a browser viewer (rubiks.html )

  • lib/shlex.ll ,lib/fnmatch.ll ,lib/textwrap.ll ,lib/csv.ll ,lib/json.ll ,lib/re.ll β€” CPython modules rewritten in llang and verified against the stdlib

  • lib/pylev.ll ,lib/humanize.ll ,lib/box2d.ll β€” external open-source repos cloned and rewritten, verified against the originals (1,562 cases);box2d is a headless physics engine subset with a working 60-step simulation

  • examples/box2d_visual.ll β€” writesphysics.html , a live browser animation of that simulation (all frames computed in llang)

  • lib/llang_lexer.ll β€” llang's own tokenizer, written in llang (self-hosting stage 2)

  • lib/advanced_math.ll ,lib/datascience.ll ,lib/machinelearning.ll β€” domain libraries grounded in Python math/statistics, NumPy/pandas and scikit-learn (330 cases)

  • 2,546 differential cases total, 0 mismatches

  • src/llang/compiler.py β€” WebAssembly backend;tools/playground/ β€” browser playground

  • src/llang/lsp.py β€” language server;src/llang/packages.py β€” package manifests and registry

  • editors/ β€” VS Code and tree-sitter grammars;docs/SITE.md β€” documentation site guide

  • tools/port.py β€” automated Python β†’ llang converter with an honest coverage report

  • src/llang/portapp/ β€” universal repository porting app; seedocs/PORTING-APP.md

  • tools/cleanup_prompt.py β€” turns a cleanup-queue item into a ready-to-paste AI prompt

  • tools/check.py β€” one command for tests, lint, rewrite verification, and link checks

src/llang/    lexer, parser, canonical AST, interpreter, synonyms, errors,
              formatter, linter, converter, WASM compiler, LSP, packages,
              CLI entry points, foreign allowlist
lib/          standard library written in llang
examples/     runnable programs
tests/        unit, differential, diagnostics, end-to-end, compiler
docs/         reference and design documentation
docs/research/ grounding documents from authoritative external sources
tools/        port.py, check.py, compare_*.py, fuzz_semantics.py, playground,
              fixtures, pilots
editors/      VS Code TextMate grammar and tree-sitter grammar
.github/      CI workflow and issue templates
run.py        development CLI entry point
repl.py       interactive shell
── more in #ai-tools 4 stories Β· sorted by recency
── more on @llang 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain β€” perfect for shipping the agent you just read about.

$git push zahid main
β†’ Live at https://your-agent.zahid.host βœ“
Get free account β†’ Pricing
from €0/mo Β· no card required
LIVE [news/natural-programming-…] indexed:0 read:6min 2026-11-22 Β· β€”