journal-adapt-writing-skill: Journal Style Adaptation

TL;DR

TL;DR: journal-adapt-writing-skill is a two-layer academic writing framework that layers a corpus-driven dynamic writing skill on top of optional static writing skills, so Claude Code or Codex can revise a manuscript to match a specific journal’s local writing culture instead of generic academic English.

Source and Accuracy Notes

• Project page: github.com/WantongC/journal-adapt-writing-skill
• Source repository: github.com/WantongC/journal-adapt-writing-skill
• License: MIT (verified via GitHub API license.spdx_id)
• Topics: academic-writing, claude, economics, journal, latex, llm, manuscript-revision, prompt-engineering, research, writing
• Source last checked: 2026-06-16

What Is journal-adapt-writing-skill?

journal-adapt-writing-skill is a revision framework for academic papers that splits the writing problem into two layers. The static layer is a reusable base writing skill such as an economics writing guide, a CS research-paper skill, a philosophy skill, or a generic AI-writing cleanup rule set. The dynamic layer is generated for one manuscript and one writing destination by reading the actual papers that define that destination. The framework outputs a dynamic_writing_skill.md next to the manuscript, plus per-section revisions, a style profile, and a revision log.

The project’s design argument is that “academic writing” is not one thing. Two journals in the same field can have different introduction logic, different contribution framing, different method exposition, different result emphasis, and different discussion scope. A revision assistant that treats all journals the same will produce revisions that read fine in general but fit no specific venue. journal-adapt-writing-skill tries to make the destination an explicit input and to extract a local writing culture from a user-provided corpus.

It is not an auto-writer. It does not generate facts, citations, results, or claims that are not already in the manuscript. It produces a visible, editable revision framework and per-section rewrites, and it expects a human reviewer to approve the result before the manuscript is sent to a real journal.

Repo-Specific Setup Workflow

Step 1: Install the skill

For Claude Code:

mkdir -p ~/.claude/skills/journal-adapt
cp -R skill/* ~/.claude/skills/journal-adapt/

For Codex, copy or symlink the skill/ folder into your Codex skills directory. You can also keep the repository open and ask Codex to use skill/SKILL.md directly.

Step 2: Prepare the corpus

The minimum Markdown workflow needs a corpus folder plus the manuscript. The PDF workflow requires a PDF-to-Markdown converter. MinerU is supported, but Markdown is the recommended path if MinerU is hard to install.

my_project/
├── corpus/
│   ├── target_journal_001.md
│   ├── target_journal_002.md
│   └── field_top_paper_001.md
└── manuscript.md

Recommended starting sizes: 5-8 target-journal papers, 2-5 secondary corpus papers, and 1-3 user or lab exemplars.

Step 3: Invoke the skill

/journal-adapt

Or ask the agent in natural language:

Help me build a dynamic writing skill for my manuscript using these target-journal papers and this base writing skill.

The skill prompts for the target journal, the primary corpus folder, the optional secondary corpus folder, the optional user or lab exemplars, an optional base writing skill, the manuscript file, and the sections to revise.

Step 4: Read the generated artifacts

Saved next to the manuscript:

[manuscript_name]_revised/
├── dynamic_writing_skill.md
├── style_profile.md
├── abstract_revised.md
├── introduction_revised.md
├── ...
├── [section]_revision_log.md
└── revision_summary.md

The revised files are Markdown. Move them into LaTeX, Word, or another writing environment after human review.

Deeper Analysis

The priority system

The framework uses an explicit priority chain to decide what to follow when sources disagree. P1 hard constraints always win, including preserved facts, citations, equations, notation, numerical results, labels, and author-defined terminology. P2 is the target journal corpus. P3 is the secondary corpus and user exemplars. P4 is the static base skill. P5 is cleanup rules that remove AI-taste phrases, hollow transitions, generic contributions, and unsupported overclaims. Conflicts that change revision behavior are recorded in the revision log so the human reviewer can audit them.

Static layer choices

The project does not try to ship a single opinionated base skill. It points at existing open-source static skills for economics, ML and CV and NLP research, CS systems, philosophy, and general AI-writing cleanup. Authors can also bring their own SKILL.md, their lab or advisor guide, or skip the static layer entirely. The dynamic corpus is designed to work even when no static base skill is selected.

Dynamic corpus

The dynamic corpus is the main feature. The primary corpus is the target journal’s own published papers, which carry the local writing culture. The secondary corpus is field-top or topic-similar papers, useful when the target journal corpus is small or when the topic needs extra reference points. User and lab exemplars are author-side preferences that should be preserved when they do not conflict with the target journal.

What the framework does not do

The README is explicit about limits. It only supports English-language academic writing. PDF conversion quality depends on the converter, and MinerU can fail on some setups. The project extracts writing structure and rhetorical patterns and must not quote or paraphrase copyrighted corpus papers. The generated dynamic skill needs human review before revision begins. The tool does not add facts, citations, results, or claims that are not already in the manuscript.

Practical Evaluation Checklist

• [ ] Do you have a clear target journal or writing destination before starting?
• [ ] Can you assemble 5-8 papers from that journal in clean Markdown?
• [ ] Do you have 2-5 secondary field-top or topic-similar papers ready?
• [ ] Have you decided whether to layer a static base skill, and which one?
• [ ] Is your manuscript already a complete draft with stable section structure?
• [ ] Are you prepared to review the generated dynamic_writing_skill.md before letting the agent revise?
• [ ] Will you audit the per-section revision logs before sending the paper anywhere?
• [ ] Do you understand that the tool will not invent facts or citations for you?

Security Notes

The skill runs locally inside your agent environment. The corpus, manuscript, and revision artifacts all stay on your disk. The risk surface is essentially the same as any other LLM-based writing tool that consumes local files. Be careful about the corpus: do not include unpublished manuscripts, confidential reviews, or copyrighted full texts that the project does not have permission to use. The README is explicit that the project must not quote or paraphrase copyrighted corpus papers, and the same caution applies to user exemplars.

If you adapt the framework for a team, treat the dynamic_writing_skill.md and style_profile.md as collaboration artifacts. They can leak unpublished work and lab-specific writing preferences, so keep them in private repositories. The framework does not call out to external services for the writing step itself, but the underlying agent may use the same providers it always uses.

FAQ

Q: Does journal-adapt-writing-skill generate a paper from scratch? A: No. It produces a corpus-derived dynamic writing skill and per-section revisions of a manuscript you already have. It does not add facts, citations, results, or claims that are not in the manuscript.

Q: Can I use it without a target journal? A: The framework is designed around a specific writing destination. If you do not have a target journal yet, the static base skill path is a better fit, or you can pick a representative venue to build the corpus from.

Q: How big does the corpus need to be? A: The README recommends 5-8 target-journal papers, 2-5 secondary corpus papers, and 1-3 user or lab exemplars. Smaller corpora work but the style profile will be noisier.

Q: Can I use the framework with Codex as well as Claude Code? A: Yes. The skill folder is designed to be installed into both. The install instructions cover the Claude Code path and notes that Codex users can keep the repository open and point the agent at skill/SKILL.md.

Q: Is the dynamic skill safe to reuse across multiple manuscripts? A: The dynamic skill is generated for one manuscript and one writing destination. It encodes the manuscript’s structure, so reusing it for an unrelated paper is not the intended workflow. Static base skills are the reusable layer.

Conclusion

journal-adapt-writing-skill is one of the more honest takes on LLM-assisted academic writing. It treats journal style as an empirical, corpus-derived property rather than a generic English grammar exercise, and it keeps the human in the loop through explicit priority rules, a visible dynamic_writing_skill.md, and per-section revision logs. The cost is real setup work: you need a clean target-journal corpus and a complete manuscript. For a researcher preparing a submission to a specific venue, that is a fair trade for revisions that actually sound like that venue.

Related reading: GitHub Trending tools, Developer tools, Adam CAD, MobileGym, ponytail.