Skip to content

julianken/language-agent

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

language-agent

An AI agent that teaches you to read a target language — Arabic (MSA) and Traditional Chinese instances, built simultaneously on one drop-in engine — from zero, through a conversation that blends English and the target language.

You talk to a capable conversational agent (it can search the web and read documents) about whatever you like. Its replies come back in a progressive English↔target blend that starts almost entirely in English and, word by word, becomes the target language as you learn — so you are always reading slightly above your level but never lost. Every target-language word is clickable, tinted by how well you know it, and sheds its scaffolding as you master it (Arabic: the short-vowel marks / ḥarakāt; Chinese: the pinyin ruby, tone last). You watch the training wheels come off, in the text itself.

Status: in design — pre-code. There is no application build, run, or test yet. The design system, screens, and the vocabulary engine are fully specified (see DESIGN.md, the Figma file in INSTANCE.md, and docs/plans/); implementation hasn't started. The repo's own scaffolding checks (CI + scripts/ guards) are live from the first commit.

The idea

  • Two decoupled axes. A reply's difficulty moves along vocabulary (which words are Arabic) and grammar (sentence structure) independently, so the blend can advance one without overwhelming the other.
  • A verified-blend engine. The learner must never see an incorrect target-language byte. Replies are generated as verified target text (vocalized MSA; Traditional Chinese with verified readings) and then anglicized down by substitution — never generated in English and machine-translated up. Deterministic gates and diverse LLM juries guard correctness end-to-end (neither the maintainer nor the operator reads either language, so verification is load-bearing, not a nicety).
  • The vocabulary system decides which words a learner sees, when. The blend ratio is the emergent sum of per-word state, capped by a coverage policy, aimed by a frequency curriculum, sequenced by the conversation topic, and advanced over spaced encounters (FSRS). The full locked design is in docs/plans/vocabulary-system.md.
  • Progress-as-typography. The reading surface is the progress display — word-status tints and the scaffold-fade (vowel-fade in Arabic, pinyin-fade in Chinese) are the core semantic system, rendered in running text rather than chrome.

How this repo is built

Most code here is written by AI coding agents under human review, then squash-merged. The process — conventions, the review-before-merge rule, agent guardrails, and the Update-Triggers table that keeps docs from drifting — is the source of truth in AGENTS.md. It was stood up from the agentic-seed template.

Map of the repo

Read these in the order START_HERE.md gives:

File / dir What it's the source of truth for
START_HERE.md Entry card — read first to orient.
INSTANCE.md Instance facts: the product, the gh slug, the Figma file + node map, the merge/review infra, lifecycle status.
AGENTS.md Process — how agents work here. (CLAUDE.md / GEMINI.md are thin shims that import it.)
DESIGN.md The design system — tokens, color, type, motion, components, accessibility. Wins on any design conflict.
docs/plans/ Committed specs & plans — the product brief, the v1/v2 design specs, the motion spec, the user stories, and the vocabulary system.
docs/decisions/ The decision log.
docs/research/ Durable research — the Arabic-NLP validation and the competitive teardown.

Security

See SECURITY.md for how to report a vulnerability and what to expect. The repo is public to support the maintainer's writing and to keep an open agentic-development trail; public status implies no compliance or audit obligation. There are no secrets or personal data in this repository.

License

MIT — see LICENSE.

About

An AI agent that teaches you to read Modern Standard Arabic through a conversation that blends English and Arabic

Resources

License

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors