ML paper writing

Make the scientific argument easy to follow: what question the paper answers, what the evidence establishes, how each step follows, and why it matters. Preserve scientific meaning, justified uncertainty, and the author's voice. The core guidance is academic writing principles. Supporting workflows are task-specific references, not mandatory steps for every prose revision.

Read only what the writing task needs

Read writing-guide.md for drafting or prose revision. For a full paper or multiple findings, also read narrative-structure.md. Use manuscript-review.md to review the result.

Load a supporting workflow only when the task needs it:

These references preserve existing project conventions and supplement, rather than replace, the core principles. A prose-only edit does not require an export pipeline, venue conversion, or submission audit.

Evidence is optional reading, not a prerequisite for ordinary editing:

• Additional oral-paper readings: seven papers from 2023–2024, with terminology, skim paths, headings, figures, and connected-passage observations.
• Original oral-paper study: 25 papers, emphasizing section architecture, transitions, and contribution scope.
• Sources: evidence limits and author preferences.

Core principles

Apply these across abstracts, main text, captions, tables, appendices, and supplementary prose. Adapt their form to empirical, theory, resource, and position papers; do not impose a fixed template.

1. Make the central question and answer recoverable early. Explain the concrete problem, relevant gap, approach, supported finding, and consequence. Do not hide the answer to create suspense or start with generic field history.
2. Support skimming without replacing the argument. The abstract stands alone; the introduction explains the reasoning and contribution; headings identify questions or supported findings; figures and captions expose the key comparisons. These layers must agree in terminology and scope. Useful repetition orients readers; repeated summaries must not crowd out explanation.
3. Order sections by questions and prerequisites. Establish a definition, measurement, assumption, or premise before a result depends on it. Keep shared setup together and analysis-specific setup beside its comparison.
4. Give each paragraph one argumentative job. State its point, develop the necessary evidence or reasoning, and end with what advances that point. A finding can span several paragraphs. Split when the purpose changes, not at a fixed sentence count. Do not put a heading over every paragraph.
5. Make transitions logical, not decorative. Within paragraphs, connect familiar context to new information and keep referents explicit. Between paragraphs, results, and sections, carry forward the established object or inference and name the dependency, contrast, unresolved question, or changed comparison axis. Parallel findings need not imply one another.
6. Use technical vocabulary only when it earns its place. Prefer a precise ordinary phrase to unnecessary jargon, acronyms, or coined labels. Explain indispensable unfamiliar terms before relying on them; define metrics and symbols at first substantive use. Use one name for one concept consistently, including headings and captions. Repetition is better than a misleading synonym.
7. Explain formal machinery. Give the purpose or intuition before an equation, definition, theorem, or algorithm and its scientific consequence afterward. Keep assumptions attached. An example supports understanding, not a proof or an untested generalization.
8. Give figures distinct argumentative roles. Order displays by the questions they answer. A conceptual overview, decisive comparison, explanation, and boundary test do different work. Merge redundant takeaways, not necessary controls or context. Preserve names, comparison order, and encodings across prose and displays; captions orient the reader, while text develops the inference.
9. Use only as much hierarchy as navigation needs. Add a heading for a substantial change of question, method component, or evidence role. Prefer a topic sentence for a short continuation. Avoid nested one-paragraph sections, repeated run-in labels, and headings that interrupt a connected finding.
10. Be precise without defensive prose. Distinguish observation, association, intervention, theorem, hypothesis, and illustration. Scope claims directly; preserve local qualifications that change their meaning. Develop broader restrictions in Limitations rather than ending every paragraph with a caveat.

Scientific and editorial invariants

• Never invent results, procedures, citations, uncertainty, or novelty. Read the primary passage supporting a citation. Keep units, denominators, comparisons, and assumptions consistent across every account of a result.
• Keep scientific procedures self-contained. Remove internal paths, run names, configuration keys, code identifiers, release bookkeeping, and development history from manuscript prose, including appendices. Preserve necessary public model/dataset names, experimental settings, and required software citations.
• Prefer analysis-generated numerical macros and .tex tables for changeable results. Reconcile prose, captions, and tables against the same result source; refreshed values still require a fresh interpretation.
• Rewrite invented compounds and overloaded modifier stacks as clear phrases or clauses. Keep established technical spelling, official names, and grammatical hyphens that prevent ambiguity; do not merely remove punctuation.
• Preserve the author's effective cadence and level of formality. Do not force conversational documentation style, second person, universal active voice, vocabulary bans, or artificial variation onto academic prose.

Text and visual design

This skill covers the argumentative role of figures and their relationship to prose. For creating or redesigning displays, read the figure-design companion. Keep the scientific question, source evidence, comparison, units, uncertainty, and scope intact. Do not turn a prose-only revision into an unsolicited graphics project.

Before delivery, apply the review checks to the requested scope. Fix the text rather than merely listing issues unless the user asked for critique. Report specific unresolved evidence or unavailable checks; do not manufacture completeness. Oral selection supplies examples, not authority over these principles or proof of writing quality.