Going further

Take a single contract clause and give it to the model three ways: plain English; then salted with Latin maxims (contra proferentem, res ipsa loquitur, volenti non fit injuria); then citation-dense and heavily abbreviated. Ask for a one-line summary of each, and watch them get shakier as the text gets denser — legal language costs more tokens and breaks into less familiar pieces. Then ask the model to estimate how many tokens a 50-page PDF would take, and how many such documents would fit a 200k-token window.

Try it in any Claude or Gemini chat · exercises your Token counter

Concatenate fifteen to twenty documents into one very long prompt — Gemini's long context is well suited to this — and bury one decisive fact in the middle: say, a 21-day deadline to accept a settlement offer. Ask a question that turns on it. Then move the same fact to the end and ask again. If the answer changes with position, you have reproduced, in your own hands, the live "lost in the middle" debate the lesson sets out.

Try it with Gemini's long context · exercises your Context budgeter

Load a folder of public sources into NotebookLM (which grounds its answers in the sources you give it and cites back to them) or a Claude Project, and slip in one document that is boiler-plate-similar to the one that actually answers your question — a decoy. Ask the question, then check which document it cites. This is Reuter et al.'s "retrieved the wrong document" failure, run on your own corpus.

Try it in NotebookLM or a Claude Project · exercises your Chunk-and-retrieve

Two halves. First, ask for "the leading authority on [a narrow point] in [a jurisdiction]", then verify every case it names against BAILII or the official reports and count the fabrications — Large Legal Fictions, reproduced by your own hand. Second, give the model five source passages and instruct it to summarise using only those, flagging anything it cannot ground; set that against the free-wheeling version and see what the discipline bought you. (Public or synthetic documents only — see rule 1.)

Try it in any Claude or Gemini chat · exercises your Grounding check

Hand the model twenty claims and one structured instruction: apply the grounding check to each and return a pass/flag table — that is fan-out by prompt. The real, agentic version is Claude Code with an MCP connector to a citation database, which turns the course's mocked citation-checker into a live one. Part-way through a long run, ask the model to restate which source each conclusion rests on: that is re-grounding, by hand.

Try it with Claude Code + an MCP connector · exercises your Batch-and-verify

Use your two models as a real bench. Ask Claude and Gemini the same evaluative question independently, then show each one the other's answer and ask it to find the weaknesses in the reasoning and the citations. Because they are trained differently, the disagreement is genuine — a diversity you cannot get from one model wearing three hats. (One-model fallback: have a single model play three named critics — one checking citations, one the statute, one the reasoning — and watch how readily they converge once they "discuss".)

Try it across Claude and Gemini together · exercises your Diverse panel

Every time you correct a model, file the correction — into a Claude Project's custom instructions, a Gemini Gem, or a Claude Code skill or CLAUDE.md. Do it for a week and you will have a working commonplace book that heads off each failure the next time it would have happened. That is the course's whole loop, running in a real tool — and the skill that makes every other skill.

File corrections into Projects, Gems, or Claude Code · this is the meta-skill