Writing research papers: rules for agents

Condensed from paper-machine. Drop this file in a repo where an agent will write or edit a paper. The full skills carry the reasoning and the incidents behind each rule.

Non-negotiable

Never write a BibTeX entry from memory. Every reference comes from Semantic Scholar, DBLP, CrossRef or arXiv, or it does not go in the file. Agents produce entries that are almost right: real authors, plausible venue, no such paper. LaTeX cannot detect this. Only a database lookup can.

Never write a number you have not just read from a file. Open the file under results/ and copy the value, even if you produced it yourself earlier in this session. Your memory of a number is a generated guess. Never write a comparison you have not computed.

Never simulate a result. If an experiment cannot run, mark it blocked and say so in the paper. "We could not evaluate X because Y" is a normal sentence in a normal paper. Invented numbers are misconduct.

Never install into a shared environment. Use a project-local venv or conda prefix. One pip install into a shared env once downgraded a core library and silently broke every user of that machine for five days.

Before writing

• Ground the paper in real prior work you retrieved and read.
• If the literature tool fails, retry, then fall back to direct academic search. If the survey is degraded, say so in the paper. A thin foundation is acceptable; an invisible one is not.

While writing

• Claims, not activities. "We conduct extensive experiments" says nothing.
• Match claim strength to evidence. One dataset is not "across domains".
• Report the experiment that did not work. Its absence is conspicuous.
• Leave an evidence trail beside numbers: % [evidence] results/exp1.json -> metrics.f1 = 0.8234

Figures

• Plot only values that exist in results/.
• The file's bytes must match its extension. A JPEG saved as .png compiles fine and then fails hard when a vision model reads it. Re-encode rather than rename, because the filename is already referenced.
• Never display a bitmap wider than its native resolution supports.
• Every figure must be referenced by \ref in the text.

Compiling

Run the full chain, every time:

pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex

A bare pdflatex re-run does not refresh the bibliography, and page-fitting passes often trigger one. Verify the delivered PDF, not an earlier one.

Before saying it is done

The paper is finished only when all of these hold:

• main.pdf exists, opens, and is over 50 KB
• Zero ? markers in the rendered text
• Every reference resolves in an academic database (state the count)
• Every number traces to a file (state the count)
• Every figure is referenced in the prose
• The AI-assistance disclosure appears in the rendered PDF

If any fails, report the failure and what would fix it. Do not report success. A truthful failure is worth more than a confident lie.