Legal AI Course

Date
2026-06-30
Status
Active — in active development
Source
github.com/richardsummerville/legal-ai-course
Demo
legal-ai-course.pages.dev

A seven-module course that teaches legal academics how to work reliably with AI on legal material. It is deliberately not a developer course: there is nothing to install, no account, no backend. Every lesson runs in the browser, and you learn by editing inputs and watching the output move. Claude Code and agentic tools are the destination it is an on-ramp to — not, yet, its subject.

Shape

The modules follow a single legal workflow as it grows from simple to complex. Each one assumes the discipline taught by the one before, so the difficulty curve is the dependency chain made visible.

ModuleDiscipline
L0FoundationsTokens, context windows, statelessness — felt, not described
L1Getting documents inContext engineering
L2Finding what mattersRetrieval / RAG
L3Summarising reliablySkills and claim grounding
L4Working at scaleParallel work and connectors
L5Reasoning acrossSubagents — a moot-court panel of perspectives
L6Improving the loopAgentic engineering

L1 through L5 each emit one artifact into a running toolkit/; L6 reflects on the loop that produced them. The aim is that a participant leaves with a working cognitive toolkit rather than a folder of notes.

Why it’s built this way

The hard constraint is that the audience are researchers, not programmers, and the material is often sensitive. So the whole thing is static and anonymous by design — no audio, no documents, and no identity ever leave the reader’s machine. In-browser Python (Pyodide, wrapped in a small edit-and-perturb cell component) makes the lessons genuinely interactive without a server to trust.

It also takes its own evidence seriously. The academic spine — the “lost in the middle” long-context limits, legal hallucination rates and their mitigations, RAG over legal corpora, multi-agent evaluation — is anchored to named primary sources, each tagged with an honest verdict (strongly evidenced / emerging / contested) rather than asserted.

Stack

Astro + Starlight for the static site and lesson prose; a React–Pyodide island for the interactive cells, with JupyterLite kept for the multi-cell lab views. Deploys to its own Cloudflare Pages project. Per-lesson frontispieces are public-domain images, sourced and attributed the same way the wiki articles on this site are.