ManuWright

He will not let you cite a paper you have not registered.

Plan first · every citation registered · every number from your data · gates before submission
_{A medical-manuscript workflow for AI agents: Claude Code, Codex, Antigravity, opencode and Muse share one verification engine.}

_{한국어 · 日本語 · 中文}

You know him. The senior author in every good lab. Reads the draft with a pen in hand. Circles a sentence and asks, quietly, "Which paper says that?" Circles a number: "Which table?" Nothing leaves the lab until he signs it.

ManuWright puts him inside your AI agent.

Before / after

You ask your agent to write the Results section before the plan is settled.

Without ManuWright, it writes fluent prose, cites a plausible paper that does not exist, and rounds a mean it never read.

With ManuWright:

BLOCKED by workflow gate (Rule 8): drafts/draft_plan.md has not been approved.
Create the draft plan first, get author approval, then draft sections.

And once the plan is approved, every claim has to survive the checkers:

$ manuwright citations drafts/03_introduction.md
GATE FAIL
citation: EVID:smith_2021
reason: citation id not found in knowledge/evidence.md

$ manuwright numbers drafts/05_results.md
GATE FAIL
number: 54.3
reason: number not found in results CSV files

How it works

 evidence.md   ┐
 results/*.csv ┼──▶ approved plan ──▶ draft ──▶ gates ──▶ QC ×3 ──▶ signed build
 style spec    ┘                                  ├ citations
                          ▲                       ├ numbers
                    hooks block                   └ semantic review (independent)
                    writes without it

The engine never invents approvals or reviews. It records decisions that humans made.

Academic writing mode

The agent writes in the register of medical journals from the first draft, not in chatbot prose. The targets were measured from 33 openly licensed original articles in JAMA Surgery, JAMA Network Open, Lancet, BMJ and Nature, published 2019 to 2022, before generative AI. Only the numbers ship (docs/academic_style/reference_profile.json); the papers stay off GitHub. For example, "used" appeared 345 times against "utilized" once, and delve, underscore and "plays a crucial role" never appeared.

You use it from the chat:

It works only inside a paper folder. Rewrites never change the facts: manuwright style preserve fails if a citation, number, p value or Table/Figure reference moved. Walkthrough: manual, section 6.

What you teach it stays. Say "from now on use used, not utilized", "remember: Spine wants a capital P" or "add Dr Kim as a co-author" and the agent saves it in your personal library (~/.manuwright/library: wording rules, notes, team profile), adding without overwriting. Every new paper starts with it, and a paper you already started pulls the new items when you open a session. Rules you approve from your own edits go there too. "Sync my library" runs manuwright library sync --pull (or --push to send a paper's rules up).

Install

Two ways, one engine. Pick one.

A. Template (no install). Clone the repository and write inside it. Claude Code picks up .claude/; Codex and Gemini read AGENTS.md and GEMINI.md.

git clone https://github.com/grotyx/Academic_writing_c_claudecode my-paper

B. Installed engine. One CLI for all your papers, plus adapters for each agent.

uv tool install git+https://github.com/grotyx/[email protected]
manuwright agents install --dry-run     # preview, then run without --dry-run
manuwright setup                        # picks the main model and reviewers for you; updates, Obsidian
manuwright init my-paper
manuwright target                      # inside the paper: target journal + Word style

Update. manuwright update installs the newest release and then refreshes the agent adapters (manuwright agents update; skip with --no-agents), then updates the agent rules of every registered paper without asking (manuwright init --refresh-rules --all --auto; only unedited AGENTS/CLAUDE/GEMINI.md copies that ManuWright wrote change, .bak kept; a file you wrote or edited is left alone, as is a paper that pins another engine version). A paper folder you open in Claude Code or Codex also updates its own rules at session start, so nothing is missed on Windows or for papers not yet registered. manuwright check shows in one screen whether everything is current (version, each agent's plugin, main model, OpenRouter key, auto-update, paper rules) and the command that fixes each ✗. Opt in to automatic patch updates with manuwright config set auto-update on: it checks once a day and never applies an update that would invalidate a paper's current review. On Windows (uv install) Windows cannot replace the running manuwright.exe, so manuwright update opens a PowerShell window that waits until ManuWright has closed, installs the new version and refreshes the agent adapters and paper rules; if no window can open, it prints the same steps as one line (uv tool install --force ...; if ($?) { manuwright agents update; manuwright init --refresh-rules --all --auto }) to paste into a new PowerShell window. Template users: git pull, or see the migration guide.

Uninstall. claude plugin uninstall manuwright@manuwright, codex plugin remove manuwright@manuwright, agy plugin uninstall manuwright, muse skills uninstall manuwright, then uv tool uninstall manuwright.

Commands

In Claude Code the template also ships slash commands: /verify, /search-evidence, /import-doi, /style-pass, /verify-claims, /suggest-citation, /cite-stance, /evidence-table, /paper-debate, /critical-review, /editor-review.

Your Obsidian library (optional, recommended)

Keep your references in Obsidian with Academic Paper Citation Manager, a plugin by the same author (Obsidian Community plugins): PubMed import, AI summaries, MeSH tags and [@citekey] citations, with your notes as the database. Its MCP server lets every agent search that library while it writes.

manuwright obsidian install          # plugin into a vault, offers MCP access (skip if installed)
manuwright obsidian connect          # register its MCP server (rag-obsidian) with each agent
manuwright evidence import-obsidian <citekey>

The vault is for discovery; only knowledge/evidence.md is citable. An imported note brings its CSL fields as the citation and the plugin's AI summary as the summary fields, is cited as [EVID:<citekey>], and stays abstract-only until someone reads the full text. Keep Obsidian open with MCP access on while agents use it. Walkthrough: manual, section 3b.

The workflow

Full rules: WORKFLOW.md. Everything that used to live in this README (features, project structure, document list): docs/guide/overview.md.

FAQ

Does it write the paper for me? It helps an agent draft inside your rules. You decide the message, approve the plans, and sign off. Human and co-author review stay mandatory.

Does anything leave my machine? Checks run locally. Sending text to another model (external review, PubMed search) happens only when you run those commands.

Where do my real manuscripts go? In a private folder or repository, never in a public clone of this template. manuwright init creates one.

Why "ManuWright"? Manuscript + wright, an old word for a maker: playwright, shipwright. Someone who builds manuscripts carefully.

Documentation

User manual · Harness guide · Workflow reference · Verification protocol · Writing guide · QC guide · Distribution design · Changelog

Author

Professor Sang-Min Park, M.D., Ph.D., Department of Orthopaedic Surgery, Seoul National University Bundang Hospital, Seoul National University College of Medicine. https://sangmin.me/

License

CC BY 4.0. Copyright (c) 2026 Sang-Min Park, Seoul National University Bundang Hospital. Share and adapt for any purpose with attribution. Full text: LICENSE. To cite ManuWright, use CITATION.cff (GitHub "Cite this repository").