research-writing

name	research-writing
description	Write papers where every claim maps to evidence and every citation resolves.
stack	sci

Research Writing

Use this skill when drafting a paper, report, or technical writeup where claims must be traceable to evidence. Structure follows IMRaD (Introduction, Methods, Results, Discussion); every sentence that asserts a fact carries a citation or an evidence marker.

When to Use

• Drafting a paper, preprint, thesis chapter, or technical report.
• Converting experiment results into a results section.
• Reviewing a draft for claim–evidence mismatches.

Procedure

1. Outline before prose

1. Write the IMRaD skeleton: Introduction (gap + contribution), Methods (reproducible), Results (measured only), Discussion (interpretation + limits).
2. Decide the claims each section must support; list them before writing so the evidence is gathered first.
3. Mark which claims are new (this work), established (cited), or uncertain (flagged).

2. Write Methods as a recipe

1. Every number in Methods is a parameter another researcher can set: versions, seeds, data splits, hyperparameters.
2. Methods must contain enough to reproduce without asking the authors — cross-check with reproducible-experiments.
3. State the statistical tests, corrections, and exclusions used (cross-check with data-analysis).

3. Keep Results evidence-bound

1. Each result sentence refers to a table, figure, or recorded measurement — no orphan claims.
2. Report negative and null results alongside positive ones; state sample sizes and uncertainty.
3. Figures and tables are generated by scripts in the study workspace, not pasted by hand.

4. Cite what you read

1. Every citation key resolves to a literature store entry with a doi or url (see literature-review).
2. Cite the primary source for a claim, not the survey that mentioned it.
3. Distinguish "X reported Y" (cited) from "we observed Y" (this work) — never blur them.
4. Verify each URL/DOI resolves with web_extract before submission.

5. Review for claim–evidence fit

1. Read the draft once with the question "which evidence proves this sentence?" — flag every sentence with no answer.
2. Check that Discussion claims do not exceed Results evidence (no overclaiming).
3. Run the spelling/consistency pass and confirm every citation key appears in the reference list and vice versa.

Pitfalls

• Writing Results before the measurements exist, then fitting prose to whatever numbers appear.
• Claims in Discussion that Results cannot support.
• Citing a secondary source when the primary paper is recorded.
• Orphan citations (in text but not in the list, or listed but never cited).
• Overclaiming: "we prove" when the evidence is a single run.

Verification

• Every claim maps to a recorded measurement, table, figure, or literature entry.
• Every citation resolves; none are orphaned.
• Methods would let another researcher reproduce the work.

Cross-references

• Use literature-review for the evidence base and citation trail.
• Use reproducible-experiments and data-analysis for the numbers behind Results.
• Use benchmark-evals when Results compare methods.
• Use project-hygiene to keep drafts, evidence, and scripts in a Project/ workspace.