Human is the first open source AI programming language that makes it possible to code using LLMs.
Requires Claude Code and uv.
uv tool install humanlang
human skills # installs the /human and /decompile skills into ~/.claude/skillsUpgrade to the latest version with:
uv tool upgrade humanlang
human skills # refresh the installed skills too
human init # run in each project to refresh its readerIn your project:
human init # creates the human/ folder: the maps and the reader live thereThen in Claude Code:
/human <abstraction>— compile free text into code, mapped to your words./decompile <file>— explain an existing file, pinned to its lines.
human serveOpen http://localhost:8010/human/web.html. Every entry opens as a notepad, and "compile" hands your words to claude. A "new file" line at the top of the file tree makes an empty file, in a new folder too; write its abstraction and claude writes the code under it.
human train --open # start a session; /decompile now writes three versions per file
human train --close # map the versions you picked and finish the sessionPick in the reader at http://localhost:8010/human/web.html: an open session shows as a layer over it, swipe sideways for the versions, down for the next file, and say why you picked if you want. A "close training" button applies the picks; a new abstraction you do not pick takes best, a sync you do not pick carries over. A write in the reader that asks claude for a telling opens a session by itself. The sessions live in the training store, one JSON per session under your user and per project, with a copy of the code and of the abstraction each row worked on; human login <key> once per machine names you, and human init gives the project its id.
Everything the project writes lives in the human/ folder — one rm -rf human/ removes it completely.