CoCalc Guides

A CoCalc-AI field guide for math grad students

From Notebook to Paper

Polish a LaTeX draft, regenerate the evidence behind a result, and keep every change reviewable with Codex, Jupyter, and project history working in the same collaborative project.

Illustrated workflow connecting a LaTeX paper, Codex review loop, and Jupyter evidence

This guide follows Maya, a second-year math graduate student turning a rough note into a paper draft. She already has the usual pieces: paper.tex, a bibliography, a few figures, and a notebook that produced one important table.

The point is not to let an agent write the mathematics. The point is to keep the paper, computations, and review loop close enough that good judgment stays cheap.

CoCalc-AI is also multiplayer. The same files, notebooks, terminals, chat, agent threads, and project history can be part of one shared session instead of being scattered across private laptops.

01

Set up the paper desk

Maya opens one CoCalc project and creates a workspace rooted at papers/spectral-gap. That workspace becomes the boundary for the files, terminals, notebooks, and Codex thread for this paper. If a collaborator joins, they see the same working surface instead of a screenshot of it later.

Working rule: one workspace, one argument, one review thread. Codex should see the same tree Maya is thinking about.
Illustration of a focused CoCalc paper workspace with files, draft, and agent thread
02

Ask for a structural read, not a rewrite

She opens the workspace Agent and asks Codex to read the draft like a careful collaborator:

Read paper.tex and tell me where the narrative breaks. Do not rewrite the paper yet. Focus on missing definitions, unsupported claims, theorem order, and places where notation changes meaning.

Codex can inspect the project and respond in the workspace thread. Maya keeps the first pass diagnostic: it becomes the checklist for the editing session.

Illustration of Codex turning a LaTeX draft into a structural review checklist
03

Regenerate the evidence

One theorem depends on a table from experiments.ipynb. Maya opens the notebook beside the paper, reruns the relevant cells, and asks the Agent to explain a failing cell instead of pasting an error into a separate chat.

When the notebook produces a new table, Codex updates the LaTeX around it: caption, label, paragraph reference, and the short explanation after the table.

Illustration of a Jupyter notebook result flowing into a LaTeX paper figure or table
04

Apply precise edits

Now Maya asks for small patches, one at a time. A good prompt names the file, the goal, and the constraint.

In paper.tex, improve the paragraph before Theorem 3.2. Keep the theorem statement unchanged. Add one sentence explaining why the numerical evidence in Table 1 supports the conjecture, but do not overclaim.

This keeps Codex in the role of editor and build assistant. The math remains Maya's responsibility; the patch is easy to inspect.

Illustration of a narrow reviewable Codex patch to a LaTeX paragraph

The useful loop

1 Ask narrowly

Name the file, the target section, and what must not change.

2 Run the evidence

Use Jupyter for the claims that depend on computation.

3 Build the paper

Let warnings and failed references become concrete tasks.

4 Review the patch

Accept only changes whose mathematical intent you can defend.

05

Review the trail

Before sending the draft to her advisor, Maya reviews the edited files and the Agent thread. Project history gives her a way to walk back through the session instead of treating AI output as a pile of anonymous text.

If the advisor is already in the project, the handoff can be live: same PDF build, same notebook result, same terminal session, same agent thread.

The final pass is deliberately boring: rebuild the PDF, check the bibliography, open every figure, and read the statements aloud.

Illustration of a review trail from rough draft through notebook run, patch, and final PDF

What changes about the work?

The paper is still written by a mathematician. The difference is that the editing loop becomes local, inspectable, and reproducible: the draft, computation, agent thread, terminal, PDF build, and project history are all part of the same collaborative workspace.

CoCalc-AI is most useful when it shortens the distance between a question and a checked change.