Academic Writing skill for Claude Code

A paper is done well when a reader can state its contribution in one sentence after the abstract, every claim is matched to evidence of the right strength, every citation exists and says what the text says it does, the figures tell the story without the prose, and the manuscript follows the venue’s format and disclosure rules.

Hard rules

• Never fabricate data, results, citations, quotes, or page numbers. Every reference must be verified to exist and to support the specific claim (read it, or at least its abstract and the relevant section). Unverified references are marked [VERIFY].
• Never state a result stronger than the analysis supports. Correlation is not described as causation; a single dataset does not “generalise”.
• The authors’ ideas, data, and conclusions stay theirs. Improve structure and prose; flag gaps in reasoning instead of inventing content to fill them.
• AI assistance must be disclosed according to the target venue’s policy (see below). Remind the user of this whenever substantial drafting help is given.
• Keep the user’s terminology, notation, and numbers consistent across sections.

Workflow and writing order

Write in the order that makes the argument stable, not the order of the paper:

- [ ] 1. One-sentence contribution and target venue
- [ ] 2. Figures and tables (the evidence)
- [ ] 3. Methods
- [ ] 4. Results
- [ ] 5. Discussion and limitations
- [ ] 6. Introduction (now you know what you are introducing)
- [ ] 7. Related work
- [ ] 8. Conclusion, abstract, title (last)
- [ ] 9. Pre-submission checks and disclosures

1. Contribution statement

Before any prose, write:

Problem: <what is unknown or broken, for whom>
Gap: <why existing work does not solve it, with 2-3 key citations>
Contribution: We <verb: show / propose / measure / prove> <what>, which <result with number>.
Evidence: <the 2-4 figures or tables that prove it>
So what: <who should change what they do because of this>

If the contribution cannot be stated in one sentence, the paper is not ready to draft. List contributions explicitly as bullets at the end of the introduction (expected in most computer science venues, useful everywhere).

2. Structure by paper type

Type	Structure	Notes
Empirical journal (IMRaD)	Introduction, Methods, Results, Discussion	Results report; Discussion interprets. Do not mix.
CS / ML conference	Introduction (with contribution bullets), Related Work, Method, Experiments, Discussion/Limitations, Conclusion	Page limits are strict; appendix for proofs and extra results
Theory paper	Introduction, Preliminaries, Main results, Proofs (or sketches plus appendix), Discussion	State theorems early; intuition before proof
Thesis	Introduction, Background/Literature, 3-5 core chapters (often adapted papers), Discussion, Conclusion	Each chapter needs a linking paragraph to the thesis argument
Short paper / letter	Compressed IMRaD; one main result	Cut related work to essentials

3. Section standards

Introduction (move structure, after Swales’ CARS model):

1. Establish the territory: why the area matters (one paragraph, no throat-clearing).
2. Establish the niche: what is missing or contested, with citations.
3. Occupy the niche: what this paper does, the main result, contribution bullets, roadmap.

Methods: enough detail to reproduce: participants or data sources with n and selection criteria, materials, procedure, analysis plan, software and versions, ethics approval and consent, preregistration if any. Link code and data or state why not.

Results: in the order of the research questions; each paragraph opens with the finding, then the statistics. Report everything planned, including null results.

Discussion: main finding in plain words, how it relates to prior work (agreement and conflict), mechanism or explanation with appropriate hedging, limitations that are real (not ritual), implications, future work that follows from the limitations.

4. Abstract formula

One sentence each, 150 to 250 words unless the venue says otherwise:

Context: why the problem matters.
Gap: what is not known or not solved.
Approach: what was done (design, data, n, method).
Results: the main finding with key numbers (effect size, CI, or benchmark delta).
Implication: what it means and for whom.

Check the venue: some require structured abstracts (Background, Methods, Results, Conclusions). No citations, undefined abbreviations, or claims absent from the paper.

5. Related work: synthesise, do not list

Listing (avoid)	Synthesis (do)
“Smith (2021) used CNNs. Lee (2022) used transformers. Park (2023) used graphs.”	“Prior approaches differ mainly in how they model long-range dependencies: convolutional models [Smith 2021] capture local structure but struggle with long contexts, while attention-based models [Lee 2022; Park 2023] scale context at quadratic cost. None evaluates on streaming input, which is our setting.”

• Organise by theme, approach, or question, not by author or year.
• Each paragraph ends by stating how this paper differs or what remains open.
• Compare fairly; describe competing work as its authors would recognise it.
• For a systematic search and synthesis matrix, use the literature-review skill.

6. Figures and tables

• Each figure makes one point; the caption states the point, then explains the encoding (“Error bars show 95% CIs across 5 seeds”).
• Captions are self-contained: define abbreviations, n, and what error bars mean (SD, SE, or CI; never leave it unstated).
• Axis labels with units; readable font size at print width; colour-blind-safe palettes; do not encode meaning by colour alone; vector formats (PDF/SVG) for plots.
• No 3D charts, dual y-axes only when unavoidable, start bar charts at zero.
• Tables: no vertical rules (booktabs style), consistent decimal places, units in headers, best results bolded only with a stated criterion and variance shown.
• Every figure and table is referenced in the text, in order.

7. Citation discipline

• Cite the primary source for a finding, not a review or a blog that mentions it.
• Cite what you read. If a claim comes from a secondary citation, say “as cited in” or find and read the original.
• Each citation must support the exact sentence it is attached to. Over-citation (“[3-17]”) hides which paper supports what.
• Self-citation only where genuinely relevant; anonymise for double-blind venues.
• For .bib cleanup, DOI resolution, and style conversion, use citation-management.

8. Reporting statistics

• Report the test, statistic, degrees of freedom, exact p, effect size, and confidence interval: t(48) = 2.31, p = .025, d = 0.65, 95% CI [0.08, 1.22].
• APA style: exact p values to two or three decimals, p < .001 below that, no leading zero for values that cannot exceed 1 (p, r), confidence intervals in square brackets.
• Lead with effect sizes and intervals; a p value alone says nothing about magnitude.
• State n per group, exclusions and why, missing-data handling, and whether analyses were preregistered or exploratory. Correct for multiple comparisons or justify not doing so.
• Distinguish statistical from practical significance in the Discussion.
• ML and benchmark results: report mean and standard deviation (or CI) over multiple seeds, the number of runs, the same tuning budget for baselines, and the evaluation split.
• If the user’s analysis lacks effect sizes or CIs, compute them from their data or ask; do not invent them.

Hedge in proportion to evidence:

Evidence	Wording
Randomised experiment, replicated	“X increases Y” / “demonstrates”
Single experiment or strong quasi-experiment	“X increased Y in our study” / “indicates”
Observational association	“X was associated with Y” / “suggests”
Speculation beyond the data	“One possible explanation is…” / “may”

9. Venue fit

• Aims and scope match; the venue has published similar papers in the last 2-3 years
• Format: word or page limit, template, reference style, anonymisation rules
• Open access requirements from the funder and article processing charges
• Indexing and legitimacy checked (Think. Check. Submit. checklist; beware unsolicited invitations and guaranteed fast acceptance)
• Cover letter: contribution, fit with the journal’s readership, no concurrent submission

AI-use disclosure

Policies differ by publisher and change; check the venue’s current author guidelines before submission. Common ground across major policies (ICMJE, ACM, IEEE, Springer Nature):

• AI tools cannot be listed as authors; humans are accountable for all content.
• Substantive use (drafting text, generating code, figures, analysis) must be disclosed.
• Where it goes differs: ACM and IEEE ask for the acknowledgments section, with IEEE asking which sections used AI and how; Springer Nature asks for the Methods (or equivalent); ICMJE asks for disclosure in the submitted work with the tool and its purpose.
• Basic grammar and spelling assistance is usually exempt, but the boundary varies.
• AI-generated images are restricted or banned at several publishers (for example Springer Nature, with narrow exceptions).
• Reviewers must not upload manuscripts to AI tools without the venue’s permission.

Draft disclosure (adjust to the venue):

During the preparation of this work the authors used <tool, version> to <purpose:
e.g. improve the readability of Sections 2 and 5 / draft initial code for the analysis
pipeline>. The authors reviewed and edited all output and take full responsibility for
the content of the publication.

Responding to reviewers

Number every comment, answer each one, quote changes with page and line numbers, and disagree with evidence. For the full process, response-letter template, and rebuttal format, read references/reviewer-response.md whenever the user has a decision letter or rebuttal to write.

Gotchas

• Writing the introduction first locks in claims the results later do not support.
• Results and Discussion blend: interpretation creeping into Results, new results appearing in Discussion.
• “Novel” and “first” invite reviewers to find a counterexample; show novelty through the gap instead.
• Inconsistent numbers between abstract, text, tables, and figures are the most common reviewer catch; check every repeated number.
• Limitations sections that list trivia while omitting the real threat to validity.
• Acronym soup: define on first use in the abstract and again in the body; cap new acronyms.
• Double-blind venues: self-references in the third person, no identifying links, scrubbed PDF metadata.

Verification

• The contribution sentence from step 1 appears, in substance, in the abstract, introduction, and conclusion.
• Every number in the abstract matches the results section and tables.
• Every reference has been checked to exist (DOI or stable URL) and to support its sentence; any [VERIFY] markers are resolved or reported to the user.
• Statistics include effect sizes and intervals; error bars are defined in every caption.
• Format, length, anonymisation, and AI-disclosure requirements of the target venue are met.

Output contract

When revising or reviewing a manuscript, return:

## Contribution (one sentence)
## Structural issues (ordered by impact)
1. <section>: <problem> → <fix>
## Claim-evidence mismatches
| Claim (location) | Evidence offered | Issue | Suggested wording |
## Statistics and figures to fix
## References to verify
## Venue and disclosure checklist status
## Revised text (if requested), with changes marked