Pick the shared page before you pick the tricks.
Teams arrive with unequal tooling comfort. Some authors live in the terminal; others have never typed a git command and never should. Forcing one tool on everyone loses authors at both ends. Hence a ladder: eleven ways to share one paper with an LLM in the loop, from zero-install to full automation. Every author picks their own rung; git (or a shared doc) merges the difference.
Two dimensions, not one. The rungs sort by two things at once: how much infrastructure you stand up (browser-only → git + CI), and how much autonomy the LLM gets (copy-paste chat → an agent that commits and opens PRs). Most rungs walk the diagonal; the interesting corners are off it — top-left is an agent loose inside a shared doc, bottom-right is a git purist who lets the LLM check grammar and nothing else.
Eleven rungs. Filled red: the four this paper actually ran. Hover a dot for its name.
The full ladder.
| 0 | shared Google Doc; the built-in AI (Gemini) drafts and comments in place. Zero install, zero accounts beyond Google. |
| 1 | shared doc + chat LLM by copy-paste (Google Docs + claude.ai; Word/Copilot and Notion AI sit here too). This paper's rung 1. |
| 2 | live LaTeX in the browser: Overleaf plus its own AI assist. Needs only a project invite. This paper's rung 2. |
| 3 | collaborative markdown pad (HackMD) + copy-paste LLM; convert to LaTeX later. |
| 4 | GitHub web UI + Copilot: prose via pull request, no clone, no terminal. |
| 5 | local clone, git, one subdirectory (or file) per person, LLM in the editor or CLI. This paper's rung 3. |
| 6 | dual remotes: the git bridge; one push updates GitHub and Overleaf together. This paper's rung 4. |
| 7 | the LLM becomes a co-author: an agent opens branches and PRs; a human merges (this repo's standing rule for agents). |
| 8 | CI on every push: build the PDF, lint the prose, publish the site. |
| 9 | CI critics: Reviewer-2 agents (trick 6) run automatically on every push. |
| 10 | scheduled agents draft and critique overnight; humans wake to PRs to accept or reject. |
Running the four rungs of this paper. Two habits make every rung mergeable: start a new line after every full stop (git then merges by the sentence), and keep prose in many small files, one per section, joined by \input under a thin root.
| rung | one-time setup | daily cycle |
|---|---|---|
| 1: Google Docs + claude.ai | open the shared Doc link | draft and comment in the Doc; paste passages to claude.ai for rewrites; paste results back as suggestions |
| 2: Overleaf browser | accept the project invite; set Menu → Settings → Main document to the paper being built | edit prose in sec0/; new line after every full stop; comment via \note{you}{...}; Cmd-Enter recompiles |
| 3: GitHub | git clone github.com/timm/how2rite | pull; edit sec0/; make fmt; commit; push |
| 4: the git bridge | rung 3, plus: in Overleaf, Account Settings → Git Integration → Generate token (username git, password the token); then git remote add overleaf https://git.overleaf.com/<project-id> | git pull overleaf main (collect browser edits); edit; make fmt; build; push to both: git push origin main; git push overleaf main:main |
Combines with: everything — step 0 is the floor the eight tricks stand on. In particular 6 — assemble critics becomes rung 9 when wired into CI, and 7 — pause and reflect is what keeps rungs 7–10 answering to humans.