Anything to Journal

Anything to Journal is an open-source Agent Skill for building a journal or conference manuscript from whatever research material you have. Put the materials for one paper in a fresh folder, invoke the skill, choose a generic draft or a specific venue, and let the agent reconcile the complete source set into an auditable manuscript project.

The workflow accepts mixed inputs: PDFs, Word files, notes, Markdown, LaTeX, spreadsheets, datasets, figures, tables, reference exports, supplementary files, code, and official publisher templates. Every input receives a stable source ID and SHA-256 record. Used evidence maps back to the manuscript; unused material receives an explicit reason.

What you get

• a journal or conference manuscript in editable LaTeX;
• a compiled PDF that has been checked page by page;
• a source manifest, material review record, traceability ledger, and quantitative evidence map;
• preserved figures, tables, equations, citations, and supporting files;
• author-decision and quality-gate reports;
• a local-only Manuscript Workspace with a scrollable PDF preview and syntax-highlighted LaTeX editor;
• clickable PDF references for mapped terms, values, figures, and tables, with source-file previews and yellow highlights;
• submission/overleaf-upload.zip, ready for New Project → Upload Project in Overleaf;
• a complete audit package for reproducibility and handoff.

The skill never invents data, references, authors, ethics approvals, permissions, or venue requirements, and it never submits a manuscript on the author's behalf.

The workflow

fresh material folder
        │
        ├─ choose: generic draft or specific venue template
        │
        ├─ inventory and review every source
        ├─ map evidence, citations, figures, tables, and equations
        ├─ write the journal manuscript
        ├─ confirm author-only decisions
        ├─ compile, inspect every PDF page, and audit
        │
        ├─ journal-output/submission/overleaf-upload.zip
        └─ local Manuscript Workspace: PDF Preview | LaTeX

The format choice is deliberately made before source content is opened. A named journal or conference requires its current official instructions or template. A generic run uses the bundled publisher-neutral interchange profile and remains clearly labeled as a draft.

Quick start

1. Create one fresh folder

my-paper-materials/
├── study-notes.md
├── methods.docx
├── results.xlsx
├── analysis.csv
├── figure-01.png
├── references.bib
└── official-template.zip       # optional

Keep all material for one manuscript together. Do not place unrelated projects or the generated journal-output/ from an earlier run in this folder.

2. Install the skill

Recommended — install with npx

Node.js 18 or newer is required. Install the latest published release with:

npx anything-to-journal@latest install

The default destination is $CODEX_HOME/skills/anything-to-journal, or ~/.codex/skills/anything-to-journal when CODEX_HOME is unset. Restart Codex only if the skill does not appear automatically.

To update an existing installation later, run:

npx anything-to-journal@latest update

install never overwrites an existing target. update first verifies that the destination is an Anything-to-Journal skill, stages the new release, and then replaces the old copy atomically.

For a repository-local installation under .agents/skills, use:

npx anything-to-journal@latest install --repo /absolute/path/to/repository

Use --destination /absolute/path/to/skills for another explicit skills directory. Run npx anything-to-journal@latest --help for all options.

Ask your agent to install it

Copy this repository URL and send it to your agent:

Please install this Agent Skill for me: https://github.com/howardtuan/Anything-to-Journal

Contributor install from a clone

Clone this repository, then run:

python3 install.py

Development installs use one canonical source through a symlink:

~/.agents/skills/anything-to-journal -> <this-repository>/skills/anything-to-journal

Use python3 install.py --mode copy for a standalone copy, or --repo /path/to/repository for a repository-local installation. The installer refuses to overwrite an existing destination.

3. Ask the agent

Open the fresh material folder, invoke the skill, and add your request after the skill name. For example:

/skill Anything-to-Journal Turn everything in this folder into a journal manuscript.

Before reading the materials, the agent asks you to choose:

• Generic draft — a publisher-neutral editable manuscript; or
• Specific venue — a named journal or conference using current official instructions or an official template you provide.

After that answer, the agent reads and classifies every file, drafts from the evidence, asks for human-only declarations and approvals, builds the PDFs, and returns the output folder.

When the environment can keep a localhost process running, the agent then starts the Manuscript Workspace and returns its http://127.0.0.1:PORT URL. In Codex Desktop it can open that exact page in the side browser when the browser capability is available; the same URL also opens in Chrome, Safari, or Edge.

Source locations are recorded in manuscript/evidence-locations.json, extending the existing evidence-map.csv. See the location format and examples. Older projects retain file-level references; missing or ambiguous locations are shown explicitly. PDF text, common UTF-8 text/CSV, DOCX extracted text, and raster image regions are supported; other formats retain an original-file download and their recorded locator. Scanned PDFs need a recorded region; there is no automatic OCR. PDF.js is bundled locally with the skill.

Direct workspace preparation

Agents normally run this for you. After the format decision, a generic intake is:

python3 skills/anything-to-journal/scripts/prepare_workspace.py \
  /absolute/path/my-paper-materials \
  --output /absolute/path/my-paper-materials/journal-output \
  --draft-only \
  --confirmed-by "REQUESTING USER" \
  --confirmation-note "User explicitly requested a generic journal draft."

A target-venue intake records the official evidence:

python3 skills/anything-to-journal/scripts/prepare_workspace.py \
  /absolute/path/my-paper-materials \
  --output /absolute/path/my-paper-materials/journal-output \
  --target-venue "VENUE NAME" \
  --venue-type journal \
  --official-guide-url "https://official.example/author-guide" \
  --guidance-file /absolute/path/my-paper-materials/official-template.zip \
  --confirmed-by "REQUESTING USER" \
  --confirmation-note "User chose the venue before source access."

Use --venue-type conference where appropriate. Inputs are copied with safe filenames and are never executed during intake.

Output anatomy

journal-output/
├── source/
│   ├── materials/                    # immutable copies
│   ├── source-manifest.json          # IDs, paths, sizes, media types, hashes
│   └── inventory.md                  # readable source index
├── manuscript/
│   ├── manuscript.tex                # main editable source
│   ├── references.bib
│   ├── traceability.csv
│   ├── evidence-map.csv
│   └── evidence-locations.json
├── reports/
│   ├── format-decision.json
│   ├── source-review.json
│   ├── author-decisions.json
│   ├── visual-inspection.json
│   ├── quality-report.md
│   └── workspace-invalidation.json  # appears after post-audit source edits
├── submission/
│   ├── overleaf-upload/              # expanded editable project
│   ├── overleaf-upload.zip           # upload this one file
│   ├── manuscript.pdf                # or DRAFT_NOT_FOR_SUBMISSION.pdf
│   └── submission-package.zip
└── project.json

The source ledger uses stable src-material-NNNN IDs. Every material ends as verified with a real output mapping or not-used with a concrete reason. Quantitative manuscript claims carry adjacent evidence markers matching evidence-map.csv.

Edit in the local Manuscript Workspace

The finished-paper loop is:

Anything-to-Journal
        ↓
generated LaTeX + PDF
        ↓
start Manuscript Workspace
        ↓
PDF Preview | LaTeX
        ↓
Codex edit or manual edit
        ↓
save and recompile the PDF preview
        ↓
formal build + page inspection + audit

Start it directly with:

python3 skills/anything-to-journal/scripts/workspace_editor.py \
  /absolute/path/journal-output

The command prints a URL such as http://127.0.0.1:43127 and stays running until stopped. --port 0 is the default and chooses a free port. Add --open-browser to ask the operating system to open the same URL in its default browser.

There is only one web UI and one set of source files:

• PDF Preview is the default tab. Its embedded PDF viewer scrolls through the complete paper, updates after every successful preview compile, and retains the previous successful PDF when compilation fails.
• Source references place clickable nodes beside mapped PDF text. For example, click the reference beside “Model accuracy was 0.9.” to see the supporting file, its page or lines, and the exact source quote highlighted in yellow. Figure/table regions can be outlined, and claims with multiple sources let you inspect each file.
• LaTeX edits the actual journal-output/manuscript/manuscript.tex. A file picker also discovers the other flat .tex and .bib sources. The lightweight editor includes LaTeX/BibTeX highlighting, line numbers, search, undo/redo, and Ctrl/Cmd+S.
• Saving writes the actual source atomically, shows Saved/Unsaved/Compiling/Compile Failed state, then compiles after a short debounce. The Recompile button runs the same preview compile on demand.
• If Codex changes a source file from chat, the running Workspace detects it, refreshes a clean editor, recompiles, and updates the preview. If the browser contains unsaved text, it shows an external-change conflict and refuses to overwrite the disk version.

The Workspace is a preview and editing layer, not a replacement for the quality gates. Any manual or Codex source edit revokes the previous submission_ready state, visual inspection, author approval hashes, promoted submission PDF, and complete-package status. After final edits, run the existing formal commands again:

python3 skills/anything-to-journal/scripts/build.py /absolute/path/journal-output
python3 skills/anything-to-journal/scripts/audit.py /absolute/path/journal-output --require-pdf

The server binds to 127.0.0.1 only. It rejects path traversal, source symlinks, cross-origin writes, and files outside the prepared project; it uses no external CDN or manuscript upload, and preview compilation uses no LaTeX shell escape. Codex Desktop and an ordinary browser simply display the same localhost page—there is no separate Desktop UI or private Codex API dependency. CLI and headless environments can skip this optional UI while retaining the complete original build, audit, package, and Overleaf workflow.

Edit in Overleaf

Upload exactly:

journal-output/submission/overleaf-upload.zip

In Overleaf:

1. choose New Project → Upload Project;
2. select overleaf-upload.zip;
3. confirm main.tex is the main document;
4. choose the compiler named in README_OVERLEAF.md;
5. select Recompile.

The ZIP has no enclosing directory: main.tex is at its root. It includes the bibliography, required .tex/.cls/.sty/.bst files, and referenced assets, while excluding private evidence, reports, compiler junk, and the final PDF. You can edit main.tex, references.bib, supporting .tex files, and figure assets directly.

See Overleaf's official Upload a project instructions for the current interface and limits.

Build and audit

After the manuscript and human decisions are complete:

python3 skills/anything-to-journal/scripts/build.py /absolute/path/journal-output
python3 skills/anything-to-journal/scripts/audit.py /absolute/path/journal-output --require-pdf

The audit checks format confirmation, complete material review, source hashes, citation and evidence mappings, figures/tables/equations, author decisions, LaTeX logs, PDF hashes, page count, visual inspection records, and archive contents. A hard failure exposes DRAFT_NOT_FOR_SUBMISSION.pdf and the blocker IDs instead of claiming readiness.

For .docx inputs, the bundled high-fidelity adapter also inventories OOXML, revisions, fields, references, media, native objects, tables, equations, notes, comments, and relationships:

python3 skills/anything-to-journal/scripts/preflight.py source-document.docx --strict

Fixed quality rules

Unless a confirmed official template imposes a stricter compatible rule, the workflow requires:

• figure captions below figures and table captions above tables;
• figures and tables beside their first substantive callout;
• 2em first-line indentation and 0pt paragraph spacing;
• at most 19 complete manuscript pages, including references;
• no unresolved placeholders, citations, references, or unsupported quantitative claims;
• page-by-page inspection after the final build;
• final submission and accountability to remain with the human authors.

Requirements

• Node.js 18 or newer for the npx installer;
• Python 3.10 or newer;
• a TeX engine: Tectonic is preferred, with XeLaTeX or LuaLaTeX supported;
• a current browser for the optional local Manuscript Workspace;
• optional Pandoc for rich DOCX semantic conversion;
• optional LibreOffice/Word, Poppler, and image tools for high-fidelity rendering and inspection.

Check the local environment with:

python3 skills/anything-to-journal/scripts/doctor.py

Development

Run the npx installer tests, synthetic skill tests, and skill validator:

npm test
python3 -m unittest discover -s tests -v
python3 /path/to/skill-creator/scripts/quick_validate.py skills/anything-to-journal

Fixtures must be synthetic. Never commit a user's unpublished sources, private data, copyrighted publisher template, or confidential results.

License and citation

Original code and documentation are available under the MIT License. User materials, generated manuscripts, publisher templates, fonts, and third-party tools retain their own licenses; see THIRD_PARTY_NOTICES.md.

Citation metadata is provided in CITATION.cff.