AI playbook for research papers
Give this file to your AI agent with your material. It asks what you need, plans the paper, drafts it, humanizes it with Penned and builds the final PDF or Word file, without retyping the humanized text.
What you need
- An AI agent that can read files and run Python scripts, such as Claude Code, Codex, Cursor or ChatGPT with its agent mode.
- A Penned API key from your API keys page. The free plan includes 50 API calls a day of up to 1,200 words each, enough for a full paper.
- Your material: a topic, the paper you are extending, or your own project, plus notes, results, data, figures and references.
How to use it
- Download the playbook and add it to your agent's chat or project folder, together with your material.
- Send a message like the one below. Add any instructions of your own.
- Answer the agent's questions: length, LaTeX or Word, PDF or .docx, template, sections, figures and tables, references. You can reply "defaults are fine" to any of them.
- Approve the plan. The agent shows the outline, what goes in each section and an estimate of how long the work will take.
- Collect the finished paper. The agent ends with a short statement of how it was built, backed by its verification report.
Follow the attached penned-research-paper-playbook.md exactly, from Part 0 to Part 9. My notes, results and references are attached. My Penned API key is in the PENNED_API_KEY environment variable. Extra instructions: [anything else, or "none"]
Why a script copies the humanized text
Every hand edit to humanized text brings AI patterns back, even a fixed comma or a smoothed transition. So the playbook has the agent write a script that copies Penned's output into the paper's structure word for word. Humanized files are saved read-only with a checksum, and a second script checks that every humanized paragraph appears in the final PDF or Word file unchanged. The agent may only claim the paper was built this way if that check passes.
How facts stay right
Penned checks each rewrite against the draft and restores changed numbers, names and claims itself. The agent then compares every section with its draft. Anything still wrong is fixed with a small, exact patch, such as a number or a citation, that the build script applies. The rest of the sentence stays as Penned wrote it, and you get the list of patches with the paper.
How the agent stays on track
Long tasks make agents forget their instructions. The playbook has the agent keep a progress file with the rules, a checklist and the next step. The agent re-reads it and the playbook before every phase and after any break, and cannot start a phase until the previous one's checks are done.
Good to know
- It works for papers on a topic, extensions of a published paper, reviews and write-ups of your own project, not only code.
- The playbook includes Penned's research-paper prompt, so the first draft is already written the way researchers write.
- Read the finished paper once before you submit it, and run your institution's similarity check with the bibliography excluded.
- For the manual route, without an agent, see the research paper guide.