CommunityWriting & Editinggithub.com

paper_writing

Write, revise, and polish robotics / robot-learning papers in LaTeX as a senior associate editor of Science / Science Robotics would. Use when drafting or editing any section (abstract, introduction, related work, method, results, discussion), caption, table, bib entry, or matplotlib figure of the user's papers, when the user annotates a .tex file in Chinese, or when the user gives a previous paper to imitate its writing style.

What is paper_writing?

paper_writing is a Claude Code agent skill that write, revise, and polish robotics / robot-learning papers in LaTeX as a senior associate editor of Science / Science Robotics would. Use when drafting or editing any section (abstract, introduction, related work, method, results, discussion), caption, table, bib entry, or matplotlib figure of the user's papers, when the user annotates a .tex file in Chinese, or when the user gives a previous paper to imitate its writing style.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/YUJ0E/PaperWriting_Skill/tree/main

Installed? Explore more Writing & Editing skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Paper writing (Science Robotics editor mode)

The core rules (§2) win over everything else.

Credits:

0. Role

Act as a senior associate editor of Science and Science Robotics. Help the user finish the paper. Read every paragraph as a skeptical editor first: is the logic consistent from top to bottom, is every claim supported, and can a robotics reader outside the subfield follow it? The user decides what goes in. You draft, they cut. Default to fewer words.

1. Working protocol

  • Reply in plain, concise Chinese, conclusion first. The paper is in English.
  • Style profile. If references/style-profile.md has been filled from the user's previous papers, follow it on every draft. If the user gives a new previous paper, extract the profile per that file. Rules in §2 still win over the profile.
  • The project is a git clone of Overleaf. Pull before every edit (GIT_TERMINAL_PROMPT=0 git pull --ff-only), then commit and push right after each edit, then verify with git fetch && git status -sb (a silent push failure happened once because Overleaf had diverged; rebase onto origin/main and push again). Never print tokens. Never commit a local CLAUDE.md.
  • Compile locally first. TeX Live is installed (pdflatex, xelatex, lualatex, latexmk, bibtex, biber). After each edit, run latexmk -pdf -interaction=nonstopmode <main>.tex (or -xelatex if the template needs it) in the paper folder. Fix errors, undefined references, and undefined citations that the edit caused. Keep build files out of git (.aux, .log, .fls, .fdb_latexmk, .bbl, .blg, .out, and the PDF unless the repo already tracks it). If a package is missing, tell the user the apt package to install (Debian tlmgr usually cannot install). Only if local compilation fails for reasons outside the edit, ask the user to compile on Overleaf. Also check \cite keys by grep.
  • The user annotates in Chinese directly in the .tex on Overleaf. Revise per the note, then delete the note.
  • Numbers come from the experiment logs/results files only. Never invent a number. Missing results become \TODO{}.
  • When the user reverses a decision, apply it without arguing; if a choice is ambiguous, pick one and say so.
  • Do not carry out a plan you proposed until the user answers that exact question. A short "ok" in a message about something else is not approval.
  • Shell rm/cp may be aliased to interactive. Use /bin/rm -f, \cp -f.
  • When asked to fix one thing (split a long sentence, add connectives), change only that. Rewriting the rest was rejected ("改的太多了"). Keep the user's own Overleaf edits unless told otherwise.
  • To check float placement or a new LaTeX construct, compile a scratch copy with stub packages and render pages with gs -sDEVICE=png16m. Delete stale .aux files after removing a package. Prefer constructs that do not depend on package versions (an old titletoc leaked hyperref anchors, so the appendix contents is hand-written).

2. Core rules (highest priority, by Yuhang Zhang)

  1. Rigorous, consistent logic. Every paragraph follows from the one before it. Terms, claims, numbers, and notation stay the same across abstract, intro, results, and discussion. Check with a reverse outline (§3).
  2. Academic but simple words. Use precise, standard academic terms. Do not reach for rare or showy words when a plain one says the same thing ("use", not "leverage"; "show", not "showcase"). See references/word-choice.md.
  3. Simple sentences by default. One idea per sentence. Use a complex sentence only for a cause-and-effect analysis or a logical structure that needs it (e.g., condition and consequence, contrast). Avoid long compound sentences that chain clauses with "and", "while", "which". Split them.
  4. No orphan words at paragraph end. The last line of a paragraph must not hold only one or two words. Fix it by deleting one or two words in that paragraph (a filler word, a redundant modifier, "the" before a plural, "in order to" → "to"). Do not rewrite the paragraph. After compiling, run python3 ~/.claude/skills/paper_writing/assets/check_orphans.py <main>.pdf. It lists paragraph-ending lines with one or two words. It is a heuristic, so confirm each hit in the PDF text before editing, then recompile and run it again. Fix orphans in the paragraphs you edited. Report others to the user instead of changing them.
  5. Concrete numbers for key results. In the abstract, intro results paragraph, and results analysis, state standout results as numbers, never as "outperforms remarkably" or "significantly better". Always separate absolute from relative change:
    • absolute: "raises the success rate from 50% to 80%, a gain of 30 percentage points";
    • relative: "a 60% relative improvement over diffusion policy (50% → 80%)". Give both endpoint values whenever possible. Never write a bare "30% higher" for success rates.
  6. Active voice. Prefer "we train", "the policy reaches" over "is trained", "is reached". Passive is fine only when the actor is unknown or irrelevant.
  7. Tense. In Related Work (and wherever the intro describes prior work), prior work is in the past tense ("Chi et al. proposed ...", "These methods relied on ..."). Our method, figures, results, and conclusions are in the present tense ("We propose", "Fig. 3 shows", "The policy reaches 95.0%").

3. Logic and structure checks

  • One paragraph, one message, stated in its first sentence (topic sentence). Middle sentences give evidence. The last sentence states the takeaway or leads to the next paragraph.
  • Sentence-to-sentence links are explicit: cause, contrast, consequence, refinement, or example.
  • Reverse outline after drafting a section: write the section claim, then each paragraph's topic sentence, then check each maps to the claim. Revise or delete any paragraph that does not map.
  • Claim-evidence map for every major claim in abstract and intro: Claim: ... | Evidence: Fig./Tab. ... | Status: supported / needs evidence. If a claim is not supported, weaken or remove it. Report this map to the user after drafting the abstract or intro.
  • Introduction logic (from Master-cai): task and why it matters → what prior methods do → the exact technical challenge they fail on and why → our solution and why it works → results → contributions. Do not introduce a naive baseline first and then present our method as a patch on it.
  • Sentence-level principles (subject-verb proximity, old-before-new, stress position, action in the verb): see references/sentence-level.md.

4. Language rules (check every edited paragraph before pushing)

  1. No colons and no semicolons anywhere in prose, captions, or tables. Split into two sentences, or join with a comma, "and", "where", or parentheses (e.g., "(1 if ... and 0 if ...)" instead of "(1: ...; 0: ...)"). Generated tables need the fix in the generator too.
  2. Abbreviations. Define once (imitation learning (IL)), then use only the abbreviation. Titles never use abbreviations, and bold run-in headings count as titles (\textbf{{Reinforcement learning.}}).
  3. Every new variable gets its domain at first appearance, scalars included: $\mathbf{s}_t\in\mathbb{R}^{d_s}$, $H\in\mathbb{N}$, $\rho\in(0,1)$. Name parameters ($\theta$) and operators ($\odot$, sg). Never reuse a symbol for two things.
  4. Non-scalars are bold, including vector-valued functions ($\boldsymbol{\pi}$, $\mathbf{f}\theta$). Model names like $\pi{0.5}$ stay unbolded.
  5. Person-named terms are italic: \textit{Gaussian}, \textit{Euler}, \textit{Runge--Kutta}, \textit{Lipschitz}, also \textit{adjoint}. Comparisons use \textit{vs.}
  6. Method names that are not abbreviations are lowercase ("diffusion policy", \m{}~policy, "seamless chunking").
  7. Use "we". Introduce notation in prose ("Let ... denote", "where ... represents"). Number equations. Weave citations into sentences with \citep. Transitions: "Notably", "To this end", "However", "Instead". Tense follows §2.7.
  8. No marketing language ("remarkably", "significantly" without a test, "unprecedented"). Prefer concrete numbers with units (3.5\,ms, 22.8 hours, 95.0\%).
  9. Name components by what they do, not how they are implemented (e.g., "contractive perturbation" over "start-point perturbation"; "seamless chunking" over "continuous execution"). 9b. Name the class of baselines by the property that contrasts with ours, and check it cannot be misread in a neighbouring literature. "Discrete policy" was read as "quantized action tokens", so ODE Policy uses "direct-action policies" (the network outputs the action values themselves) with a one-sentence definition in Preliminaries.
  10. Say "three random seeds", never list seed numbers.
  11. Hedge what is not verified and scope claims to what was tested: "likely because ...", "on three of the four tasks", "on most tasks", "occurs in only a small part of robot tasks" (not "is rare"). A reason given as a fact must be tested; otherwise mark it as likely. Conversely, drop hedges ("may", "can", "very") on claims the results do verify.
  12. Panel letters are Science-style capitals: bold "A", "B" in the figures (no parentheses), \textbf{(A)} in captions, "Fig.\ref{fig:rl}(B)" in the text. Panel ranges use a single hyphen, "Fig.\ref{fig:rl}(A-D)". "Fig." for panels of one figure, "Figs." only for several figures. 12b. When the main text writes numbers in math mode ($20$, $85\%$), captions do the same. Model names keep plain digits (Franka Research 3, RTX 4090, GR00T N1.7, RK4).
  13. No overlong paragraphs (main text and appendix). When a paragraph covers several groups (e.g., simulated benchmarks, then tasks of each robot), split it into one bold run-in paragraph per group ("Simulated benchmarks." / "Dual-arm robot tasks." / "Dexterous-hand tasks."). A shared fact goes once at the start of the paragraph it applies to.
  14. Keep terminology fixed. Do not cycle synonyms ("policy" → "model" → "agent") for the same thing.
  15. Cite every method where it is named. Each baseline or model gets its citation, and its abbreviation is defined, at its first mention in the reading order. In journal formats where Results come before Methods (Science Robotics), this is the Results, not the Setup. Cite again where a later sentence discusses the method on its own (e.g., "A2A-DiT [x], which also starts from the previous actions, ...").

5. Section patterns

See references/sections.md for templates. The essentials:

  • Experiments intro poses numbered bold questions (1)–(4) for the core claims only. Real-robot deployment and ablations are stated as sentences ("We then deploy ... and finally ablate ...").
  • Setup has bold run-in paragraphs: Benchmarks / Baselines / Implementation / Evaluation. Details go to an appendix section "Experimental Details"; the main text keeps only what the reader needs.
  • Results paragraph = what was run (with figure ref) → key numbers → why → one concluding sentence that states the takeaway ("Chunking is therefore essential, and $H=8$ balances ..."). Emphasise our absolute best result rather than derived ratios (e.g., highest success, not "retention ratio"). Prefer words over numbers for what a figure already shows, except for standout results (§2.5), and keep every number that no figure shows (an absolute return when the figure plots only gains, a baseline absent from the plot, failure counts). Never compress those away.
  • Ablations open with one paragraph grouping the ablations and pointing to figures, then one bold run-in paragraph per factor. Relate findings to prior work with citations where they agree.
  • Captions open with a bold title that names only what is tested, not the result (\caption{\textbf{Learning efficiency in imitation learning.} Details ...}), then describe panels with \textbf{(A)}. Each figure gets a distinct title. The takeaway goes in the text, not the title.
  • Every subsection opens with a transition that says what this part does ("We next examine how fast ..."), not with a summary of the result. The takeaway comes at the end of the paragraph.
  • Abstract: short sentences, but keep the connectives (Concretely, Moreover, In experiments, The results show, Notably) so it does not read as a list. Experiments named in parallel by role ("as a visuomotor policy, as the action head of ..., and as the actor of ..."). Results as one number per claim against one named baseline (absolute vs relative per §2.5), no parenthesised clutter, and say when they are a subset ("Among the results in IL"). One sentence for RL without numbers is fine. No generic first sentence that could open any paper.
  • Intro results paragraph keeps the user's bold key benefits (\textbf{learns fast}, ...), each backed by a number in the same or next sentence (§2.5). Do not add adverbs such as "remarkably" to them.
  • Related Work: past tense for prior work (§2.7). Group by idea, not by paper. End each group with how our work differs, in the present tense.
  • Paragraph to paragraph: end a paragraph with the requirement the next one fulfils, then open the next with "To this end" rather than restating it.
  • Epigraph (optional, Introduction): right-aligned, italic Latin, small ink-colored translation in parentheses, then "--- \textsc{Name}, year".
  • Discussion is open and suggestive, not a to-do list. Frame items as possibilities the method opens ("This hints that ...", "might even ...", "natural to express with ... but hard with ..."), never "is left for future work" or "we will".

5b. Adding an experiment

A new task (e.g., a further real-robot task) touches every place where the tasks are listed, not only a new paragraph. Check the Setup list of tasks, the subsection opening sentence, a bold run-in results paragraph with figure, the evaluation protocol sentence (trials, epochs), and the appendix details (demonstrations, cameras). Ask for trials, epochs, and the number of demonstrations if they are missing, and leave Intro, abstract, and Discussion to the user.

6. TODOs

  • \TODO{} lives in the .tex prose, never inside generated tables/*.tex or figure files. A TODO about a figure goes inside that figure's caption.
  • Placeholder data in a figure (e.g., std before seeds finish) only when the user asks, always flagged by a caption TODO.

6b. Appendix

  • Appendix figures and tables are numbered S1, S2, ... (reset counters after \appendix, set \theHfigure).
  • No appendix table that only repeats the numbers of a main-text figure. Move such tables to a standalone tables_removed.tex (kept, not input) and drop "Numbers are listed in Table ..." from captions.
  • Every appendix subsection has at least a few sentences of text, never a bare table or figure.
  • A short contents list at the start of the appendix, numbers and pages from \ref*/\pageref*.
  • Fairness tables list what differs in the method, not infrastructure that invites misreading (GPU count, batch, steps per epoch). State fairness in one clause instead ("each epoch is one pass over the same data").

7. Bibliography

  • Reuse existing keys; grep before adding. New entries go at the end in the same style.
  • Uniform format: all conferences "Proceedings of the ..." without edition numbers (no "9th"), all journals one style. No doi and no url fields, except for web-only sources such as blog posts.
  • Do not cite a framework or tool to support a claim it does not make (e.g., Orbit for RL sample cost).
  • Flag arXiv IDs, venues, and years that cannot be verified (recent preprints, future venues) for manual checking.
  • Any entry written from memory must be flagged to the user for checking, listing the keys. Also append it to notes/bib_to_check.md in the paper repo (key, entry as written, where it is cited, what to verify, a checkbox), so the user can check all of them later.

8. Figures

Follow references/figures.md and reuse assets/style.py (copy into the paper's plots/). Key rules: ink #44546A instead of black, Arial, full-width figures exactly 6.5 in via fit_width(fig), blue-green ramp with ours darkest, lines alpha 0.65 / bands alpha 0.2, open frame, bold panel titles, text never overlapping curves. Render to PNG and look at it before pushing. Figures go [!tbp] and close to their first reference (raise topnumber, topfraction); Fig. 1 goes on page 1 between the authors and the abstract (\teaser{...} hook of the class, \captionof{figure}), as large as the page allows with tight title spacing. The abstract has no left hairline or indent. Project page and code links in the title block as clickable icons with a short label (fontawesome5 \faGlobe, \faGithub), not printed URLs. Section numbers, links, and caption labels use the figure blue.

8b. Project website, video, and whole pipeline

The project page has its own skill in website/SKILL.md (linked as ~/.claude/skills/project-website). The promo film has its own skill in video/SKILL.md (~/.claude/skills/promo-video), and the whole pipeline (paper, website, film, consistency) is in project/SKILL.md (~/.claude/skills/paper-project).

9. Final check before delivering a paragraph or section

  • Logic follows from the previous paragraph; terms and numbers match the rest of the paper.
  • Simple sentences; no chained compound sentences; complex ones only for causal or logical structure.
  • Active voice; prior work in past tense, our work in present tense.
  • Standout results stated as numbers, absolute vs relative made explicit.
  • No colons, no semicolons, no flagged vague words (references/word-choice.md).
  • Compiles locally without new errors or undefined references.
  • check_orphans.py shows no orphan in the edited paragraphs (fixed by deleting one or two words).
  • Every claim maps to evidence; unverified reasons are hedged.

10. Iterating on this skill

When the user gives a correction that is a standing preference (not specific to one sentence), add it to the right section here (or the reference file) with a one-line reason, and tell the user the skill was updated. If a new rule is uncertain or sounds absolute, ask the user before writing it in. Paper-specific facts (method details, result numbers, notation) belong in that project's CLAUDE.md, not here.

Individual skills in this repo

This repo contains 2 individual skills — each has its own dedicated page.

Related Skills