Deep Drafter

This page helps academics use AI to write research papers. It shares Deep Drafter, an AI agent I have developed to write papers with Claude. In 2026 I published my first philosophy paper written with AI, "Epistocracy and the Commitment Problem," in Philosophy & Public Affairs; the agent grew out of that process.

The agent is free to use and adapt: github.com/simondgoldstein/deep-drafter.

Overview

Deep Drafter allows you to write with AI the way a heavy-handed supervisor writes with a PhD student in some fields. You supply the thesis, the central arguments, and the judgment calls. Claude does the execution: literature review, outlines, drafts, revisions. You comment in detail at several stages throughout the process, including an initial plan, a paragraph-level outline, and the paper draft. You choose between implementations of the main ideas, press objections, and calibrate the strength of the thesis.

This agent is a Claude Code workspace disciplined by a set of documents: a step-by-step pipeline, a writing guide, style modules for each type of paper, and four specialist subagents. The drafting process begins with outlining. The agent builds a skeleton, then an argument outline, then a paragraph-by-paragraph outline at roughly a third of the final length, and writes no prose until you approve that outline. The user gives detailed comments at the planning, outline, and drafting stage of the process. The agent also does several rounds of adversarial blind review. The agent also has several auditors that check for errors, references, and execution of the main tasks.

Getting started

You need Claude Code and a folder. Because the agent uses a lot of tokens, it will only work well if you have a Claude Max account. Download the agent, open the folder in Claude Code, and say: set up Deep Drafter for me. Claude interviews you about your field, your venues, and your preferences, collects three to six of your best papers, and mines them for the sentence and paragraph patterns it will imitate when drafting for you. Setup takes one session. After that, starting a paper is a one-page brief and a short intake.

The download includes three example papers written with the agent: two philosophy papers and a law review article.

Failure modes

AIs writing long papers fail in characteristic ways:

  1. Organization. AIs have trouble coherently structuring long papers.
  2. Redundancy. AIs repeat the same points.
  3. Vagueness. AIs describe an argument without making it.
  4. Jargon. AIs use obscure, idiosyncratic terms.

What works

  1. Draft in many steps. I have AIs build three rounds of outlines before writing a full paper. This addresses organization, redundancy, and vagueness.
  2. Give style examples. Keep a folder of your favorite prose for the AI to imitate, and a document stating your style principles.
  3. Give comments. The more work you put into commenting on intermediate AI work product, the better the results.
  4. Request self-critique. On hard problems, have the AI generate several candidate solutions and critique them itself before you read anything.
  5. Start with an outline. I often start the process by giving the AI a 1-2 page outline of all of the central arguments in the paper.
  6. Use agents, not chatbots. Write with Claude Code or Codex, with the AI managing a folder of documents. Agents perform long tasks more reliably, and pipelines can structure their behavior.
  7. Go interdisciplinary. AIs are strongest as bridges into fields you know less well.
  8. Transparency. Be clear about how you use AI in paper writing, especially when you have new, successful methods. This helps more people learn how to use AI. It is easy to have your AI agent generate a report at the end of the project explaining exactly who did what.

Questions and improvements are welcome: open an issue on GitHub, or email me at simon.d.goldstein@gmail.com.