Skip to content

EXAMPLES.md

Latest commit

 

History

History
98 lines (76 loc) · 3.86 KB

EXAMPLES.md

File metadata and controls

98 lines (76 loc) · 3.86 KB

Examples

Concrete things people actually do with this server. The prompts below work in any MCP-compatible client (Claude Desktop, Claude Code, claude.ai web with the HTTP transport).

The pattern is the same across all of them: you describe what you want in natural language, Claude picks the right tools and arguments, the server does the work, and the change shows up in Overleaf within a few seconds.

1. Tighten the introduction against the abstract

Useful when the abstract has been revised but the intro still has stale framing.

Read the abstract and the introduction in main.tex and tell me where the introduction's framing diverges from what the abstract promises. If you find drift, propose a tightened intro that lands the abstract's claims and rewrite the introduction in place.

What happens under the hood:

  • get_sections: orient on what's in main.tex
  • get_section_content × 2: pull the abstract and the introduction
  • (Claude reasons + drafts)
  • edit_file: write the tightened intro back, full pull → write → commit → push

The commit shows up in Overleaf with the message claude: edit main.tex. Refresh the project, see the diff, accept or revert in Overleaf's history view.

2. Add a new subsection in the right place

Add a new subsection titled "Threat model" right after the "Related work" section in main.tex. Sketch the content as three paragraphs covering: assumptions about the deployment context, what we protect against, and what we explicitly don't.

What happens:

  • get_sections: find the line range of "Related work" so the insertion point is right
  • read_file: read the surrounding context to keep voice consistent
  • edit_file: write main.tex with the new subsection inserted at the right line

3. What's in this project, and what's its current state?

Useful for context-setting when picking up a project after time away.

Use the overleaf project_status and list_files tools, then summarise: how many files do I have, when was my last commit, what files are .tex vs .bib vs other, and what's the structure of main.tex?

What happens:

  • project_status: file count, dirty state, last commit summary
  • list_files: full file listing
  • get_sections on main.tex: section structure

You get back a paragraph orienting you on the project: "You have 12 files (8 .tex, 2 .bib, 2 figures), last commit was 'add methodology section' 3 days ago, and main.tex has 5 sections with subsections under Method and Results."

4. Sync from Overleaf and find what changed

Useful when you've been editing in the Overleaf web UI and want Claude to catch up before you continue working.

Sync the project from Overleaf, then read main.tex and summarise what changed in the introduction since the previous version we discussed.

What happens:

  • sync: pull latest from Overleaf, report whether HEAD moved
  • read_file: current intro
  • (Claude compares against what it had in conversation context)

Combine with project_status to see the new HEAD's commit message ("ah, you added the motivation paragraph in the web UI").

Notes on prompting

  • Trust Claude to pick the tools. You don't need to say "use get_sections first." Claude reads the tool list and picks appropriately. Telling it the intent works better than telling it the steps.
  • Be specific about what you want changed. "Tighten the introduction" is vague; "tighten the introduction so it lands the abstract's claims about X and Y" gives Claude enough to do useful work.
  • Use sync at the start of a session if you've been editing in Overleaf since the last time Claude touched the project. The edit_file tool always pulls before writing, but sync lets you see the changes before continuing.
  • Every change is a separate commit. If Claude makes a mistake, git revert in Overleaf's history view (or your local clone) is one click.