Kid Language Learning Assistant

Problem

Early-childhood language learning often falls into a pattern of parent anxiety and child resistance: words today, texts tomorrow, weekend cramming, without a stable review rhythm, individual level tracking, or visible feedback. Plain chat models can struggle to keep a child profile, progress state, overdue review items, and daily task boundaries consistent over time.

How It Works

The skill organizes family language tutoring as a file-backed workflow rather than a free-form lesson generator.
- Profile and state: On first use, it collects age, language level, and preferences, then writes child-profile.md. Later sessions read that profile first and avoid re-asking basic information.
- Storage backends: It supports IMA and local file-system backends. IMA is useful for cross-device persistence of memory files, while local is suitable for workspace debugging, scripts, and generated pages.
- Daily loop: The flow moves through Hook, Load, Action, Commit, Notify, and Maintenance. It refreshes review lists and health checks, reads today-review and daily-log, runs bilingual teaching, then writes progress back and archives logs.
- H5 output: The model emits constrained JSON only. html-injector.py injects that data into assets/template-shell.html, producing a single-file H5 page. English and Chinese days use separate templates to keep subjects isolated.
- Optional push: If a WeCom webhook is configured, the skill can push an image card or text fallback after Commit, plus an HTML file. Push failures do not block the main teaching flow.

Boundaries

It is better suited to long-term, low-pressure family language support than crash-course generation. The skill depends on local scripts, reference docs, and optional external services; if IMA, webhook, or templates are unavailable, it should fall back to plain-text plans or local data. The constraints matter: child interest first, one subject per day, bounded review capacity, and no direct HTML/CSS/JS output from the model.