Digital China Initiative · Harvard University

AI Agents for Sinitic Texts

2026 DCI Summer Workshop
Dates May 18–21, 2026
Location Harvard University
Instructor Kwok-leong Tang
Audience East Asian humanities researchers
About

About the workshop

A four-day intensive on building research tools with AI agents — for scholars working with Chinese, Vietnamese, Japanese, and Korean materials. The workshop walks from "what is an LLM" all the way to a personal LLM wiki, a published browser-based research tool, and a hands-on integration of classical digital-humanities methods (TEI, NLP, GIS, network analysis) with agentic AI.

The deliverable each participant leaves with is not a certificate — it is a tool, a wiki, and the diagnostic mindset to keep building both.

Sessions

Workshop sessions

01

Introduction and Logistics

Day 1 · Monday, May 18 · 09:30–10:30
Workshop overview, the teaching team, setup checks (LM Studio + Qwen3.5 0.8B), and demos of what participants have built with Codex.
▶ Open slides →
02

What is an LLM?

Day 1 · Monday, May 18 · 10:45–12:00
Nature and limitations of LLMs — tokens, prediction, hallucination, chain-of-thought, autoregressive behavior, and fixing what prompts cannot fix with a tool. Hands-on with Qwen3.5 0.8B in LM Studio.
▶ Reading → PDF → Slides (HTML) → Slides (reveal.js) →
03

AI Coding Agents — Codex

Day 1 · Monday, May 18 · 13:00–14:15
From a chat window to an AI agent that can read files, run commands, and act on its own. Folder-system primer; first agent task; permission modes and settings.
▶ Reading → PDF →
04

Deterministic vs Non-Deterministic — Two Ways to Extract Names

Day 1 · Monday, May 18 · 14:30–16:00
The same research task done with a regex and through an LLM API. Comparing reliability, reproducibility, and when each wins.
▶ Reading → PDF → Materials zip →
05

Finish the API loop, install Git, ship one tool three times

Day 2 · Tuesday, May 19 · 09:30–10:45
Build a single-file HTML LLM tool and publish it to GitHub Pages in three stages: LM Studio only, then add Google AI Studio and OpenRouter as cloud backends, then add image OCR. The edit → commit → push → deploy loop walked three times in one session.
▶ Reading → PDF → Materials zip →
06

Memory, Context, and the LLM Wiki

Day 2 · Tuesday, May 19 · 11:00–12:15
Why models forget, and what we put back. The two dominant strategies — Retrieval-Augmented Generation (RAG, via Google NotebookLM) and the LLM Wiki — compared. Write your first wiki page with Codex.
▶ Reading → PDF → Materials zip →
07

Obsidian and the LLM Wiki You Already Have

Day 2 · Tuesday, May 19 · 13:30–14:45
Open a real published LLM Wiki (kltng/digital-china-wiki, ~1000 pages) in Obsidian. Wikilinks, backlinks, graph view, and the four canonical Codex-on-wiki operations: read · improve · query · lint.
▶ Reading → PDF → Materials zip →
08

Build Your Own LLM Wiki — From Raw Sources to Skills

Day 2 · Tuesday, May 19 · 15:00–16:30
Scaffold a personal wiki from 29 Tang–Song fiscal-history sources (the liangsui corpus). Two-prompt workflow: scaffold, then bulk-ingest into source / people / concept / institution / debate pages. First Skill enriches one page with external data.
▶ Reading → PDF → Corpus (liangsui.7z) →
09

Enriching Your Wiki with External Data — RESTful APIs and Skills

Day 3 · Wednesday, May 20 · morning
One workflow, many backends — applied to the other API humanists meet daily: the RESTful data API. Scrape a real webpage when there is no API, then call CBDB, CHGIS, Harvard LibraryCloud, and Wikidata as portable, OpenAI-style requests wrapped in Skills.
▶ Reading → PDF →
10

Classical DH Methods, Driven by Agents — TEI, NLP, GIS, Network Analysis

Day 3 · Wednesday, May 20 · afternoon
The four durable families of digital humanities — TEI encoding, NLP, historical GIS, network analysis — are not obsolete. They are the verifiable substrate the agent runs on. The agent is the front door; the classical method is the building you walk into.
▶ Reading → PDF →
Cite & Reuse

How to cite

These materials are archived on Zenodo. A single PDF collecting all ten sessions is available there. If you use or adapt them, please cite:

DOI 10.5281/zenodo.20673640
Tang, Kwok-leong. (2026). AI Agents for Sinitic Texts: 2026 Digital China Initiative Summer Workshop. Harvard University. Zenodo. https://doi.org/10.5281/zenodo.20673640 
@misc{tang2026sinitic,
  author    = {Tang, Kwok-leong},
  title     = {AI Agents for Sinitic Texts: 2026 Digital China
               Initiative Summer Workshop},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20673640},
  url       = {https://doi.org/10.5281/zenodo.20673640}
}

Statement on AI use

These teaching materials were prepared with the assistance of AI coding agents — Claude Code (Anthropic), Codex (OpenAI), and Opencode — used for drafting, code examples, page layout, and editing. All content was reviewed, tested, and edited by the author, who takes full responsibility for it.