← all tricks

Trick 4: Write to a grammar

Every writer starts from a tested structure, not a blank page.

The move. A research paper parses under a small grammar, written in logic-programming notation: nonterminals rewrite to ordered sequences; each bracketed leaf is one prose move, a sentence to a short paragraph long.

paper --> intro, litreview, methods,
          results, discussion, conclusion.

results --> [result]+.
result  --> [rq], [charts], [comments].

intro --> hook, gap, elevator, move,
          rqs, contribs, [caveat]*, [roadmap]*.
gap   --> [problem], [naive_fails],
          [prior_fails], [key_idea].

Per-section rules, each with a worked sample:

Intro: hook = quantified pain, an anchor, or a setting; the gap is the Stanford four clauses (problem, naive fails, prior fails, key idea); the elevator is a 2–3 line epigram; each RQ's answer follows at once, headline number bold; the repro URL always closes the contributions.

Lit review: respect, then disrespect. Say clearly what prior work did well; then name the mismatch precisely (wrong assumptions, wrong data, wrong question). End with the pivot: “open and urgent issues” — and every later section must trace to that list, or be cut.

Methods: duty = replication. Inventory tables early; one worked example threads the section; every hyperparameter listed, justified (by arithmetic, budget, or prior result), and ablated.

Threats: Cook & Campbell's four categories (construct, internal, external, conclusion). A threat named without a response reads as an unfixed bug: every leaf ends with the mitigation applied or the cost knowingly accepted.

Conclusion: retells the intro, now with receipts. Nothing new. If intro and conclusion disagree, one of them is lying; fix it.

See them run. The paper pairs each grammar (left panel) with a generated-then-rewritten sample (right panel). Click to enlarge:

Introduction grammar beside its worked samplethe introduction Literature review grammar beside its worked samplethe literature review Methods grammar beside its worked samplethe methods Threats-to-validity grammar beside its worked samplethreats to validity

Why grammars beat templates. To disagree with a structure, edit its productions and ask an LLM to regenerate the worked sample; the experiment costs minutes. Your venue opens with related work, bans roadmaps, demands structured abstracts? Adjust the grammar, not your patience. When a grammar changes, regenerate its sample in the same commit so figure and prose never drift.

Call to action. Communities curate datasets and baselines; they could curate paper shapes the same way — a catalog of grammars for the empirical study, the systematic review, the tool paper, the vision paper.

Combines with: 5 — un-slop the prose polishes what the grammar drafts; 6 — assemble critics attacks the result; 8 — keep the receipts is the conclusion rule enforced.

← all tricks