A source-led research-paper workflow with Obsidian, Claude Code, LaTeX, and Codex
Disclosure: AI tools helped draft and edit this article. A human publisher reviewed the final text and workflow before publication. Putting four tools in a list does not make a workflow. The difficult part is deciding which tool owns each artifact, what the next tool may change, and what must still be checked by a person. This recipe uses four bounded roles: Obsidian keeps the source ledger,…
This article details a research paper workflow that utilizes four AI tools: Obsidian, Claude Code, LaTeX, and Codex. The process begins with the creation of a brief.md file which outlines the research question, audience, scope, deadline, rules, citation style, and any exclusions or specialist reviews required.
Next, the source ledger (source-ledger.md) is established to serve as the control surface for all material claims. Each claim is given a unique ID, the claim itself, the corresponding citekey, page/section location, status, and any additional notes. The author is responsible for updating the status of each claim from UNVERIFIED to VERIFIED after reviewing the cited source material.
Once the ledger and outline (outline.md) have been populated, the drafting agent is tasked with inventorying only the permitted source files, building the ledger, and drafting the argument in outline.md before any manual editing. This agent is instructed not to browse, download, upload, publish, or contact anyone without explicit authorization by the author.
After drafting, a build script (verify.sh) is run to check that all required artifacts exist and to build the manuscript using latexmk. A final check is then performed to reject any remaining TODO, UNVERIFIED, and UNRESOLVED markers.
Finally, a separate criticism model (Codex) is employed to review the brief, source ledger, manuscript diff, bibliography, and build output. This model identifies unsupported claims, broken references, logical gaps, and reproducibility failures, which are then documented in review-findings.md. The author is still responsible for opening each cited source and resolving any material findings.
The paper is considered ready for author review when the manuscript compiles without errors, all build warnings have been addressed, every material claim has been verified, the review log shows each finding and its resolution, and the author has confirmed the venue's authorship and AI-use rules. If proprietary tools are not suitable, alternatives such as Zettlr, Logseq, Aider, or OpenHands can be tested, although the same fixture and final checks should be rerun after making any changes.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
