Writing research papers

A paper is your work, retold so a stranger can trust it in ten minutes. Structure does half the
job; honest storytelling does the rest.

The canonical sections (in this order)

1. Title + Abstract — 5 sentences: question → data → method → headline result → what it's
   for. Write it LAST, place it FIRST. Most readers read only this; make it complete.
2. Introduction / Problem — the decision this supports, and why it matters. One real
   scenario beats three abstract paragraphs.
3. Data — source, size, time windows, what you excluded and why. Public-safe: no private
   names, addresses, or raw queries, ever.
4. Methodology — assumptions, features, how the label is defined, the baseline, the
   validation design (grouped/time split), and the leakage checks you ran. This section is what
   separates a paper from a blog post.
5. Results — model vs baseline, same split, base rate visible. Tables for precision,
   charts for shape.
6. Limitations & honest framing — what this work cannot claim. Writing this section
   yourself, before a reader does, is what earns trust.
7. Ranked recommendations — the "so what": what a person should do first, with reasons.
8. Reproducibility — links to the notebooks and repo; seeds; how to rerun.
9. Acknowledgments & data credit — name and link your data source. Standard practice, and
   it tells readers the data is real.

Storytelling that isn't spin

• One thread: question → tension ("the obvious rule fails because…") → resolution (your result).
• Concrete before abstract: open sections with the example, then the principle.
• Every chart earns its place: one message per chart, axis labels a stranger understands, the
  takeaway written UNDER the chart in one sentence. Three great charts beat ten fillers.
• Numbers in context: "24% → 68% (≈3×)" — always the comparison, never a lone number.

Reading design (people skim)

Short paragraphs (2-4 sentences). Real headings. A reader scrolling fast should learn your
finding from headings + chart captions alone. Mobile-check it — half your readers are on phones.

Repurpose the same work

• Social post: one finding + one chart + one method sentence + link.
• Employer 3-sentencer: what I built · on what data (size! realness!) · what it showed.
• SEO framing: the page title should be the QUESTION people search, not your project name.

How to verify

• The 10-minute test: someone reads only abstract + charts + limitations. Can they state your
  finding and its limits correctly? Ask your assistant to role-play exactly that reader.
• Every claim on the page passes the claim ladder (see writing-honest-claims skill).
• All 9 sections present; the data credit links; the notebooks link resolves.