🌭
Menu

Flagship workflow · editorial, not benchmarked

A reproducible AI workflow for research papers

Use Claude Code, Obsidian, LaTeX, and Codex as four bounded roles. The goal is not automatic authorship; it is a manuscript whose claims, sources, build, and review trail you can inspect.

Run the free checklistClone the starter ↗Check an artifact

The division of labor

Obsidian
Keep the source ledger, claim status, notes, and open questions in readable Markdown.
Claude Code
Work across the outline and manuscript files while every factual claim remains tied to the ledger.
LaTeX
Turn the paper, citations, equations, and build warnings into a versionable publishing artifact.
Codex
Review the source ledger, diff, and build log as a separate critic; it does not replace opening the sources yourself.

Start with this folder, not a blank chat

paper-project/
├── sources/              # permitted PDFs, links, and notes
├── source-ledger.md      # claim → citekey → verification status
├── outline.md
├── manuscript.tex
├── library.bib
├── build.log
└── review-findings.md

The source ledger is the control surface. For every material claim, record the citekey, the passage or page that supports it, whether you personally opened it, and what remains uncertain.

Use explicit handoffs

  1. Sources → outline: give Claude Code only permitted source material and require citekeys on factual statements.
  2. Outline → manuscript: preserve unsupported claims as visible questions; do not let fluent prose erase uncertainty.
  3. Manuscript → build: keep LaTeX warnings and broken references in the artifact instead of cleaning only the PDF.
  4. Build → review: give Codex the ledger, manuscript diff, and build log; ask for unsupported claims, logical gaps, and reproducibility failures.
  5. Review → decision: open every cited source yourself and resolve each finding before submission.

What “done” means

A compiled manuscript, clean bibliography links, retained build warnings, a source ledger with no unresolved material claims, and a review log showing what the second model caught and how you resolved it.

Important boundary

Models can fabricate citations, misread sources, and agree on the same error. Follow journal and institution authorship rules, do not upload restricted material, and use qualified human review where the consequences warrant it.

Primary setup references