# Thomas D. Grant

- Researcher: Thomas D. Grant
- Role: International law academic, University of Cambridge (affiliation as given in the conference programme)
- Field: law (international law)
- Tools: Codex; saved Markdown instruction files (AGENTS.md and others); a source wiki on Andrej Karpathy's pattern
- Published: 2026-11-20
- Source: https://www.arsinq.com/inferpedia/static/arsinquirendi/slides/grant-using-llms-in-scholarship-2026.pdf
- Written with the current flagship models (Claude Opus 5.5 and GPT-6.1 Sol) available.

## What they did

Grant works on international law, and his slides show his work in Codex. The programme abstract says he has been experimenting with prompts, context management and harness-building for seven months across a range of projects. The slides list seven kinds of work: brainstorming outlines, horizon-scanning the literature, identifying material imperfections in the sources, gathering pinpoint references for footnotes, proofing for consistency, finding connections in sources, and drafting abstracts. A line repeated on those slides says: "The researcher defines the task, inspects the result, and decides what to do next."

His standing instructions are kept in seven Markdown files in five functions: repository rules (AGENTS.md), his own writing conventions (TomStyle.md), instructions for each project (project.md), a prose charter, a prose shape guide and reviewer instructions for drafting and review, and a document-ingestion guide. Each new project gets eight starter files, created empty and filled as he decides the details: project.md records what was agreed (purpose, scope, audience, form), questions.md the open questions and source gaps, and sources.md the material found and what has been checked; the others are notes.md, outline.md, draft.md, authority_sketches.md and runs/current.md. His example is a lecture outline on the Non-Proliferation Treaty set up on 7 October 2026, where one sources.md entry is marked "identity and contents checked only" and records that the outline does not rely on it. The next day he told the model to tidy the working files, and the slide comments that the model's apparatus needs periodic tidying.

He also keeps a source wiki on Andrej Karpathy's "LLM Wiki" pattern: software prepares searchable entries from his source documents, each pointing back to a passage; the model examines passages and drafts a note; he assesses the evidence and decides what to keep. In one demonstration the wiki suggested a link between H. L. A. Hart and Alan Turing at Bletchley Park; reading the original passage and its endnote showed that the source supports an institutional connection only, and the note records that personal collaboration is not established by that source. He divides a piece of writing into passes with one purpose each: draft, check sources, revise, review. When a fact was missing, the drafting brief told the agent to insert a "[CITATION QUERY: ...]" marker and keep drafting, without searching for the fact.

## What they learned

Grant stresses that saving an instruction in a file and supplying it to the model are separate steps, so a file's existence does not put its contents in front of the model. When he corrects a sentence, he and the model record the general fault as a rule with examples. On 27 May 2026 a draft said "Status problems ask who counts as the sovereign"; his comment was that problems do not ask, the agent then found and repaired the same fault in another passage, and the rule was saved. By 14 August the rules file had 44 rules in 384 lines and he split it into six files by category, and he says a correction is worth saving as a rule only if it is stable, reusable and about exposition, not about one project's substance.

He reports that in one case a "[CITATION QUERY: ...]" marker was filled by a later check, and that in another the first draft cited two pages of a judgment for a claim about Article 55 of the ICSID Convention without marking any gap, until a source-checking pass found that the cited pages do not mention Article 55. His comment: "The instruction to mark gaps did not ensure that every gap was marked." Other slides record that a search memo said of its empty results "These are not findings that no commentary exists"; that the model reported when it had only a later court's summary of a source and asked for the original; that its completion report warned that two-column PDF extraction had interleaved text, so some wiki records were noisy; and that a cross-article consistency check produced findings the slides call proposals, awaiting his decision. Slides headed "a possibility to ponder" offer his conjectures, for example that adding more instructions makes it less likely the model follows each one, and that fluent writing can hide emptiness, which he proposes to counter with a separate source-checking pass.

## Limits of this account

This is Grant's own teaching deck about his own projects, and he selected the excerpts. The slides themselves mark some requests as not recovered and some results as not rerun, and they give no measure of time saved or error rates. I read all 136 slides and have not used the workshop session or the Markdown files the slides say accompany them. I have not checked his legal examples.
