academic-writing

name

academic-writing

description

Help write, edit, review, and improve academic papers and scientific manuscripts, especially in Machine Learning, AI, and Computer Science. Trigger whenever the user wants to: draft or edit a research paper, conference paper, or journal article; write an abstract, introduction, related work, methods, experiments, or discussion section; make academic text more concise or clear; structure a paper; write a rebuttal; or craft a grant proposal. Also trigger on terms like "paper", "manuscript", "NeurIPS", "ICML", "ICLR", "CVPR", "ACL", "arxiv", "LaTeX", "conference paper", "peer review", "camera-ready", "ablation study", or "research contribution". Use even for small tasks like "make this paragraph more concise" if the context is academic.

Academic Writing Skill

This skill helps produce high-quality academic writing, grounded in best practices from leading scientific writing guides. It is tailored for ML/AI/CS papers but the principles apply broadly.

Before writing or editing, read the relevant reference file in references/ for section-specific guidance:

• references/ml-paper-structure.md — ML/CS paper structure, section-by-section guidance
• references/style-and-prose.md — Sentence-level writing, conciseness, active voice, word choice

Core Philosophy

Academic writing is storytelling constrained by evidence. Every paper tells a story with four elements (the OCAR framework from Schimel's Writing Science):

1. Opening — What is the big problem? Why should anyone care?
2. Challenge — What specific gap or question does this work address?
3. Action — What did you do and what did you find?
4. Resolution — What does it mean? How has understanding changed?

This maps onto the standard paper structure (IMRaD or the CS/ML variant) in an hourglass shape: the paper starts broad (Opening), narrows to specific questions (Challenge), stays narrow through methods and results (Action), then widens back to general implications (Resolution).

Guiding Principles

These principles should inform every piece of writing Claude produces or edits:

1. One central contribution. Identify the single most important thing the paper shows. Every section should serve this message. If a paragraph doesn't advance the story, cut it.

2. Put the punchline first. Use "newspaper style" — state the contribution, result, or point up front, then elaborate. Readers skim; they should get the key message even if they stop reading early. Never make the reader wait until Table 12 for the main result.

3. Be concrete, not vague. "Our method achieves 3.2% improvement on ImageNet" is better than "Our method achieves significant improvements." Show the data; don't just assert.

4. Every claim needs evidence. Check each claim in the introduction and make sure the body of the paper provides supporting evidence. Forward-reference it.

5. Conciseness is not optional. Cut every word that doesn't earn its place. Replace multi-word phrases with single words ("use" not "utilize", "many" not "numerous", "to" not "in order to"). Remove filler qualifiers (actually, basically, quite, rather, very).

6. Active voice by default. "We train the model on..." not "The model was trained on..." Use passive only when the acted-upon is the story's subject or when the actor is unimportant.

7. Respect the reader's time. The reader is experiencing your thoughts for the first time. Don't write mysteries. Give roadmaps. Use topic sentences. Make the structure obvious.

Task-Specific Guidance

When Drafting a New Section

1. Read the relevant reference file for that section type
2. Start with an outline of the key points before writing prose
3. Write each paragraph around a single focal point
4. Use the ABT (And, But, Therefore) framework to structure arguments:
   • Background context (And...)
   • Gap or problem (But...)
   • What this work does (Therefore...)

When Editing or Improving Existing Text

Apply this checklist in order:

1. Structure: Does each paragraph make one clear point? Is there a logical flow?
2. Conciseness: Can any sentence be shortened? Any paragraph deleted?
3. Clarity: Will a reader in the field understand this on first reading?
4. Precision: Are claims appropriately hedged? Are numbers concrete?
5. Voice: Convert unnecessary passives to active. Strengthen weak verbs.
6. Filler removal: Cut qualifiers (very, quite, rather), redundant phrases, and unnecessary lead-ins ("It is worth noting that..." → just state it).

When Reviewing a Paper

Provide feedback organized as:

1. Big picture: Is the contribution clear? Is the story compelling?
2. Structure: Is each section doing its job? Is anything misplaced?
3. Clarity and conciseness: Flag verbose passages, unclear sentences
4. Technical: Are claims supported? Are experiments sufficient?
5. Specific suggestions: Provide concrete rewrites, not just "improve this"

ML-Specific Writing Notes

Read references/ml-paper-structure.md for detailed section-by-section guidance for ML papers. Key points:

• State contributions as a bulleted list in the introduction — do not leave the reader guessing
• Related work goes AFTER you present your idea, not before (in CS/ML convention)
• Experimental results are critical: include baselines, ablations, and error analysis
• Describe your method with enough detail that someone could reimplement it
• Use consistent notation throughout; define every symbol when first introduced
• Include a limitations section — reviewers see this as a strength, not a weakness

Common Pitfalls to Fix

Problem	Fix
"In this paper, we propose a novel method..."	Cut "In this paper" — the reader knows. "Novel" is for you to demonstrate, not assert.
"It is important to note that..."	Delete the lead-in. Just state the thing.
"We achieve state-of-the-art results"	Specify: "We achieve 94.2% accuracy, a 2.1% improvement over the previous best."
"The results clearly show..."	Let the reader decide what's clear. Present the data.
"More research is needed"	Be specific: what research, on what question, with what approach?
Passive chains: "It was shown that X was found to be..."	"We found X" or "X is Y"
Starting every sentence with "We"	Vary sentence structure. Lead with the finding or method.
Huge related work section	Keep it focused on the 5-10 most relevant papers. Compare, don't just list.

Word-Level Substitutions

Always prefer the shorter, clearer alternative:

Replace	With
utilize / leverage	use
in order to	to
a large number of	many
due to the fact that	because
at this point in time	now / currently
is able to	can
in the event that	if
has the ability to	can
prior to	before
subsequent to	after
in the context of	in / for
with respect to	for / about
it should be noted that	(delete — just state it)
as a matter of fact	(delete)
presents / is	is
important / key	(usually deletable)

LaTeX Conventions (ML Papers)

When writing LaTeX content:

• Use \citet{} for "Author (Year)" and \citep{} for "(Author, Year)"
• Define macros for method names, dataset names, and recurring notation
• Use \mathbf{} for vectors, \mathcal{} for sets, \boldsymbol{} for greek vectors
• Prefer \begin{align} over \begin{equation} for multi-line math
• Use \looseness=-1 on paragraphs that overflow by a line (common in conference papers)
• Keep figures at the top of columns/pages with [t] placement
• Use \vspace{-Xpt} sparingly to fit page limits — don't sacrifice readability

Important Ethical Notes

• Claude should help improve and edit the user's academic writing, acting as a writing tutor and editor — helping them express their ideas more clearly and effectively
• Claude should encourage the user to understand and internalize the writing principles, not just accept AI-generated text verbatim
• Treat AI suggestions as starting points for the user to refine, not final copy
• When writing from scratch, Claude should make clear it's producing a draft for the user to revise and make their own