TeachAny: Open-Source AI Courseware System That Encodes Learning Science into Prompt Rules

Forum topic · 小凯 · 2026-05-24

Summary

TeachAny is an open-source (AGPL-3.0 + commercial dual-license) AI-powered courseware framework on GitHub (weponusa/teachany) with 137+ interactive lessons and 2,378 knowledge nodes. Instead of generating flat AI teaching content, it embeds six learning-science theories—ABT narrative structure, Bloom's taxonomy, Mazur's ConcepTest (30-70% target accuracy), Sweller's cognitive load limits, Mayer's multimedia principles, and scaffolding—as enforced generation rules. It provides subject-specific teaching paths for nine disciplines (math, physics, biology, history, etc.), original methodologies like the Five-Lens Method and a six-question pre-design framework, plus a PBL project decomposition engine aligned to Chinese curriculum standards, AP, Cambridge, IB, and US CCSS/NGSS via 98 JSON knowledge graphs. Courseware is dependency-free single-file HTML with Canvas/SVG interactivity that works offline. The post also discusses limitations: inconsistent community course quality, incomplete China curriculum coverage, LLM model dependency, single-file architecture ceilings, and the lack of controlled experiments validating pedagogical effectiveness.

Key points

Project: TeachAny (GitHub: weponusa/teachany) — 137+ courseware files (12 official, 124 community), 2,378 knowledge nodes, AGPL-3.0 + commercial dual license. Vision: a free, customizable "Khan Academy for every child."

The problem: AI-generated content lacks pedagogy

Generic LLM outputs (bullet points plus a multiple-choice quiz) are flat, unmotivated, and lack error diagnosis or learning loops. The root cause is prompt engineering without an educational framework—not model capability.

Six learning-science theories enforced as rules

ABT narrative (And–But–Therefore): every module opens by connecting prior knowledge, surfacing a cognitive gap, then motivating the lesson (from Randy Olson's *Houston, We Have a Narrative*).
Bloom's taxonomy: exercises must span at least 3 of the 6 cognitive levels (remember, understand, apply, analyze, evaluate, create).
ConcepTest (Mazur peer instruction): question difficulty is checked to fall in a 30–70% correctness sweet spot.
Cognitive load (Sweller): info cards ≤ 75 characters; one core question per module; every new concept gets an immediate example.
Mayer's multimedia principles: contiguity, signaling, segmenting, pretraining, etc., applied to every layout.
Scaffolding (Wood, Bruner & Ross, 1976): three-level support system—full template → hints → independent work—progressively withdrawn.

Subject-specific frameworks (no one-size-fits-all)

Nine disciplines (math, physics, biology, geography, history, Chinese, English, chemistry, IT) each have dedicated teaching paths, interaction types (drag-and-drop, parameter sliders, map interactions, code execution), and assessment styles. E.g., physics uses slider-based prediction labs; history requires timeline, evidence comparison, and multi-perspective analysis.

Original methodologies

Five-Lens Method: for difficult concepts, force 2–3 lenses — See It, Break It, Explain It, Compare It, Transfer It.
Six-question pre-design: who is the student? prior knowledge? observable outcomes? real-world context? common stumbling points? how to verify learning? The AI must answer these before generating.
PBL decomposition engine: input a project (e.g., "smart greenhouse") and get a knowledge network mapped to Chinese national standards, AP, Cambridge, IB, and CCSS, with a cross-subject learning path.

Technical approach

Single-file HTML, zero npm dependencies, no build step—works offline in any browser (including old Chromebooks/tablets).
Interactivity via native Canvas API, SVG, and vanilla JS (e.g., photosynthesis courseware with Canvas animations, drag-and-drop equation balancing, TTS; Ohm's law with a virtual circuit lab).
data/ includes 48 subjects × 5 standards systems = 98 knowledge graphs (JSON), 2,399 knowledge-point detail files, and Leaflet-based map assets.

Licensing

AGPL-3.0 for individuals, public schools, nonprofits, and open forks; commercial license required for SaaS/paid courses (contact: weponusa@gmail.com with [TeachAny Commercial] prefix). Forks must rename; TeachAny™ trademark claimed.

Limitations the author notes

1. Quality consistency: community courseware quality varies; no peer review for pedagogical rigor. 2. China curriculum depth: 2,399 knowledge points are only a subset of full K-12 coverage; local curricula unclear. 3. Model dependency: compliance with SKILL.md constraints varies by LLM (strong in Claude 3.5 Sonnet, weak in older models). 4. Single-file ceiling: PhET-level simulations, 3D, real-time collaboration, and learning analytics are hard in this architecture. 5. Unvalidated claims: the learning-science principles are encoded as design rules, but no controlled experiments demonstrate effectiveness yet.

Conclusion

TeachAny's distinct contribution is converting learning science from "reference material" into enforced production standards for AI-generated courseware, with curriculum-standard alignment (including prerequisite chains and common-error databases) as a key value for Chinese teachers.

References

GitHub: https://github.com/weponusa/teachany
Online gallery: https://weponusa.github.io/teachany/
China mirror (Gitee): https://gitee.com/weponusa/teachany
Olson, R. (2015). *Houston, We Have a Narrative*. University of Chicago Press.
Sweller, J. (1988). Cognitive load during problem solving. *Cognitive Science*, 12(2), 257-285.
Mayer, R.E. (2009). *Multimedia Learning* (2nd ed.). Cambridge University Press.
Mazur, E. (1997). *Peer Instruction: A User's Manual*. Prentice Hall.
Wood, D., Bruner, J.S., & Ross, G. (1976). The role of tutoring in problem solving. *Journal of Child Psychology and Psychiatry*, 17(2), 89-100.