Research Writing Kit

What it is

Not a template. A method.

You own the ideas and the argument; the AI does the legwork: finds and verifies sources, builds the argument skeleton, writes the prose, then checks its own work. Five principles run through the whole thing:

1. Skeleton firstDon't squeeze prose out of a blank page. Text written that way tends to read like AI prose. Build the argument structure first, then write to it.
2. Every citation checked against its sourceThe method requires each reference to be fetched and its direction of support confirmed. One that cannot be found is marked "unverified" instead of being kept quietly. Before delivery it also scans for retractions and for claims with no citation at all.
3. Papers sound like your field, letters sound like youPapers, grant proposals and applications are measured against published papers in your field, so each language feature sits inside the range real papers use: no generic AI prose, and no personal idiolect either. Your own voice is kept for letters and personal statements. English drafts get a separate rhythm pass before delivery.
4. Honesty built inEffect sizes with confidence intervals, not just p-values. Likert items are not treated as continuous. Before claiming an effect, the design is checked for whether it can answer the question at all.
5. Data stays local where it canThe bundled check tools and the statistics run on your machine. Unpublished drafts and raw research data are not sent to online search services or AI detectors, and online lookups send only bibliographic data such as DOIs and titles. Anything Claude Code reads, drafts and data included, is still processed on Anthropic's model servers, so when you review someone else's submission for a journal or conference, by default it only hands you the commands for the local tools to run yourself, and reads the manuscript only if you say so and the venue allows it.

Getting started

Three steps, five minutes

1. Get the folder. It's a public repository, no login needed. Clone it with git, or use Code → Download ZIP on the GitHub page.

   git clone https://github.com/chenweichiang/research-writing-kit.git

2. Open Claude Code inside that folder. It needs a paid Claude plan; a free claude.ai account won't work.

   cd research-writing-kit
   claude

3. Paste one line and let it take over. It will ask what you write, in which language, where you submit, whether you can collect published papers from your field as a baseline, and whether you have letters of your own it can learn your voice from. Then it generates skills that are yours. "Read CLAUDE.md and set me up."

Still on the claude.ai website with no Claude Code? Connect the repo to the web Claude and say the same line. It walks you onto Claude Code step by step (setup/WEB.md).

Day to day

Just say it

Open Claude Code in the folder that holds your paper and say what you need, the way you'd say it to a person. You can also type /skill-name directly.

"Help me write this paper / proposal"/co-author

Collaborative writing: sources, skeleton, prose, and six pre-delivery gates (retractions, uncited claims, numbers ledger, figure provenance, submission statements, final review).

"This reads like a translation"/co-author

Native polish: first measures which language features sit outside the range of same-field papers; the editor subagent returns a change list only, and you rule on each item. Changed numbers, citations or terms are blocked.

"Check this draft before I submit"/paper-review

Five layers: mechanical, diction, language, reviewer's-eye logic, delivery completeness. Checks only; it does not rewrite.

"Do these citations really support what I wrote?"/verify-citations

Reads the cited PDFs sentence by sentence, judges support direction, verifies DOIs, and sends flagged items to a sceptical second reviewer.

"Run everything before I send it"/co-author

Full polish: citations, proofreading, native polish, de-AI, overclaims, localization, regression, then a new versioned PDF, with one report at the end.

"My co-author wrote this draft"/co-author

The original stays untouched and your changes become an edit layer; every sentence of the new draft is traced back to the original and the co-author's replies, and untraceable methods or numbers are flagged red.

"Get me the PDFs of these references"/fetch-refs

Multi-source fetching, per-paper content check, filing. What it cannot get is listed for you to download, minus what your own library already holds, and your downloads are filed afterwards. Includes citation snowballing.

"The reviews are back, help me respond"/rebuttal

Split the points, rule on each, make the revisions, write the letter, verify completeness before sending.

"Don't make that mistake again"/doc-regress

Turns a caught error into a standing check that scans the whole document and blocks recurrence. Includes the numbers ledger.

"Build the submission PDF"/build-pdf

Formats to the venue template; second-language papers ship with the back-translation as a pair.

"Can I submit this? What is missing?"/submit

Re-reads the official page, converts the deadline to your timezone, and leaves the scoring to a clean reviewer; builds the Word file, diffs the format against the template, checks anonymity, and after you send it files the confirmation, updates the ledger and tags the version.

What's inside

Method, templates, forty-two tools

• The method in eleven documents: philosophy, iron rules, the nine-phase workflow, argumentation moves, method decision, thirty-eight method cards, the rigour process, version numbering, co-authored manuscripts, cutting to a page limit, publishers' AI-use rules.
• Eight skill templates and five subagents (including the clean final review and native polish in both languages): adapted to your field, language, and venues at install time.
• Forty-two local scripts, most with zero install: Taiwan-usage and Chinese AI-style checks, register profiles and native-polish checks in both languages, English grammar and AI-fingerprint diagnostics, metadiscourse and lexical-bundle diagnostics, figure colour-vision accessibility and a figure check on the typeset PDF, reference retrieval and the literature map, the missing-references list and filing of hand downloads, retraction scan, uncited-claims scan, document regression, format gates for the method decision and the analysis plan, sentence-by-sentence tracing of a co-authored rewrite, comparison of a Word file a co-author returns, a second-coder pack with agreement scores, a version check, review-round convergence, response-letter and submission-ledger checks, and for the last mile deadline conversion, a Word round-trip and word count, a format diff against the official template, and an anonymization check.
• Blank templates (voice profile for letters, venue notes, argument skeleton), open academic word lists (two bundled; the AWL is fetched by a script because of its licence), one worked skeleton example.
• Traditional Chinese (Taiwan) add-on: mainland-vs-Taiwan term check, 台/臺 consistency, a Typst recipe for Chinese PDFs.

Everything in detail is in the repo's README.en.md: what appears on your computer after install, what grows inside your paper project, and which external program or online service each feature uses.

Origin and licence

Grown from one person's practice

The method comes from how Chenwei Chiang writes papers, proposals, and reviewer responses; it was de-personalized and published so researchers and students can make it their own. Code is MIT; the method and documents are CC BY 4.0, so you may adapt, translate, and teach it as long as you credit the source. The third-party word lists carry their own licences (AWL is CC BY-NC-ND); see NOTICE.md in the repo.