ViviDoc generates self-contained interactive HTML documents from a single topic input. It designs a purpose-built visual style, structures the document using the SRTC Interaction Spec (State · Render · Transition · Constraint), and writes a single .html file with KaTeX math and Canvas visualizations — no build step, no server. Your AI coding agent (Claude Code, Codex, etc.) is the harness — no external API key required.
Accepted at ACL 2026 System Demonstrations — arXiv:2603.01912
curl -sSL https://raw.githubusercontent.com/MisterBrookT/vividoc/main/install.sh | bashInstalls /vividoc, /vividoc-learn, and /vividoc-slides into ~/.claude/commands/. Your harness is the model — no separate API key needed.
/vividoc Fourier Transform
/vividoc-learn https://ncase.me/trust/
/vividoc-slides https://example.edu/lecture.pdf
| Document | Domain | Interaction |
|---|---|---|
| Fourier Transform | Physics & Math | Temporal Control |
| Lorenz Attractor | Physics & Math | Parameter Exploration |
| Action Potential | Biology | Temporal Control |
| DNA Replication | Biology | Temporal Control |
| Gradient Descent | Machine Learning | Direct Manipulation |
| Bias–Variance Tradeoff | Machine Learning | Parameter Exploration |
| Shannon Entropy | Information Theory | Parameter Exploration |
| Huffman Coding | Information Theory | Freeform Construction |
ViviDoc Learn (concept)
An email-based personalized learning product built on top of ViviDoc's generative visualization engine:
- Assess — a short interactive quiz at signup determines the learner's current level on a chosen topic
- Generate — ViviDoc produces a personalized interactive document (or short quiz) for each learning unit
- Deliver — content is emailed on a daily/weekly cadence, in the learner's preferred language
- Test — each delivery includes a small embedded quiz; results feed the next generation cycle
- Progress — staged curriculum that adapts to demonstrated understanding over time
This would be a separate hosted product ("ViviDoc Learn") from the current open-source harness skill, requiring server-side generation, email infrastructure, and subscription accounts. The current harness-skill approach remains the core open-source offering.
uv sync --dev
uv run pytest
cd frontend && npm install && npm run devBaselines (AutoGen, CAMEL, MetaGPT, naive): uv run python benchmark/run.py --baseline autogen
@article{tang2026vividoc,
title = {{ViviDoc}: Generating Interactive Documents through Human-Agent Collaboration},
author = {Tang, Yinghao and Xie, Yupeng and Feng, Yingchaojie and
Lan, Tingfeng and Lao, Jiale and Cheng, Yue and Chen, Wei},
journal = {arXiv preprint arXiv:2603.27991},
year = {2026},
url = {https://arxiv.org/abs/2603.27991}
}
@inproceedings{tang2026demonstrating,
title = {Demonstrating {ViviDoc}: Generating Interactive Documents through Human-Agent Collaboration},
author = {Tang, Yinghao and Xie, Yupeng and Feng, Yingchaojie and Lan, Tingfeng and Chen, Wei},
booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics: System Demonstrations},
year = {2026},
url = {https://arxiv.org/abs/2603.01912}
}MIT License · ACL 2026 System Demonstrations · arXiv:2603.27991