Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Review and polish LaTeX research papers for clarity and style
.claude/skills/brycewang-stanford-paper-polish-guide/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | 157% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 119% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 194% | 0% |
| case-25 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-04 | ✓→✓ | = Same ✓ | 94% | 0% |
The difference between a good paper and a great paper often lies in the final polishing stage. After the scientific content is solid and the structure is sound, a thorough polish pass addresses clarity, conciseness, consistency, and correctness at the sentence and paragraph level. This is where awkward phrasing is smoothed, redundancies are eliminated, and the prose is tightened to convey ideas with maximum precision and minimum friction.
This skill provides a systematic approach to polishing research papers, with a particular focus on LaTeX manuscripts. It covers common writing issues in academic papers, a multi-pass review strategy, specific patterns to search for and fix, and tools that can assist the process. The techniques apply across all academic disciplines but include LaTeX-specific advice for formatting and typographic polish.
The skill is designed for the final stages of writing, after the content and structure are complete. It is not about rewriting sections—it is about refining the existing prose to its best possible form.
Polish your paper in multiple focused passes rather than trying to fix everything at once. Each pass targets a specific aspect of the writing:
Read each paragraph and check:
Read each sentence and check:
Review word choice throughout:
For LaTeX manuscripts, check these formatting details:
latex% Non-breaking spaces before citations and references as shown in prior work~\cite{smith2024} in Figure~\ref{fig:architecture} in Table~\ref{tab:results} in Section~\ref{sec:methods} % Correct dash usage pages 1--10 % en-dash for ranges state-of-the-art % hyphen for compound adjectives --- and % em-dash for parenthetical (or use --) % Math mode consistency $n$-dimensional % variable in math mode even in text $O(n \log n)$ % Big-O in math mode % Consistent use of \emph{} vs \textit{} % Use \emph{} for emphasis, \textit{} only when italic is the content (e.g., species names)
\cite{} keys resolve to entries in your .bib file| Wordy | Concise | |-------|---------| | "In order to" | "To" | | "Due to the fact that" | "Because" | | "A large number of" | "Many" | | "In the event that" | "If" | | "At the present time" | "Now" / "Currently" | | "It is possible that" | "May" / "Could" | | "Has the ability to" | "Can" | | "With regard to" | "Regarding" / "About" | | "For the purpose of" | "To" / "For" | | "On the basis of" | "Based on" |
Avoid starting sentences with:
Academic writing traditionally uses passive voice, but overuse makes prose sluggish. Use active voice for your contributions and passive for standard procedures:
→ Active: "Our method improved results by..."
Complement manual review with these tools:
chktex or lacheck to catch common LaTeX issues (missing tildes, inconsistent quotes, etc.)aspell or your editor's built-in spell checker with a custom dictionary for technical termstextlint or writegood can flag passive voice, weasel words, and other style issuesbash# Find common issues with chktex chktex -v3 main.tex # Check spelling with aspell aspell --mode=tex check main.tex # Find inconsistent terminology grep -n "data set\|dataset" main.tex grep -n "pre-train\|pretrain" main.tex
Before declaring the paper polished:
\cite, \ref, and \eqref commands$3.2\,\text{ms}$| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 28,272 | 32,215 | +14% | 1 | 1 | 0% | 5,696 | 7,044 | +24% | 0 | 0 | — |
case-02 | fail→fail | 18,350 | 10,608 | -42% | 1 | 1 | 0% | 2,464 | 3,636 | +48% | 0 | 0 | — |
case-03 | fail→fail | 28,242 | 22,590 | -20% | 1 | 1 | 0% | 5,006 | 4,463 | -11% | 0 | 0 | — |
case-04 | pass→pass | 11,759 | 10,713 | -9% | 1 | 1 | 0% | 1,316 | 2,547 | +94% | 0 | 0 | — |
case-05 | pass→pass | 11,623 | 5,447 | -53% | 1 | 1 | 0% | 1,165 | 2,784 | +139% | 0 | 0 | — |
case-06 | pass→pass | 9,014 | 8,229 | -9% | 1 | 1 | 0% | 712 | 2,294 | +222% | 0 | 0 | — |
case-07 | pass→pass | 6,793 | 5,542 | -18% | 1 | 1 | 0% | 1,256 | 2,745 | +119% | 0 | 0 | — |
case-08 | pass→pass | 15,007 | 9,687 | -35% | 1 | 1 | 0% | 1,687 | 2,573 | +53% | 0 | 0 | — |
case-09 | pass→pass | 11,004 | 3,164 | -71% | 1 | 1 | 0% | 873 | 2,324 | +166% | 0 | 0 | — |
case-10 | pass→pass | 4,990 | 4,481 | -10% | 1 | 1 | 0% | 922 | 2,418 | +162% | 0 | 0 | — |
case-11 | pass→pass | 10,747 | 9,757 | -9% | 1 | 1 | 0% | 1,001 | 2,514 | +151% | 0 | 0 | — |
case-12 | fail→pass | 5,046 | 9,222 | +83% | 1 | 1 | 0% | 965 | 2,480 | +157% | 0 | 0 | — |
case-13 | pass→pass | 5,935 | 8,253 | +39% | 1 | 1 | 0% | 1,093 | 2,258 | +107% | 0 | 0 | — |
case-14 | pass→pass | 7,523 | 10,118 | +34% | 1 | 1 | 0% | 1,295 | 2,621 | +102% | 0 | 0 | — |
case-15 | fail→pass | 12,778 | 11,340 | -11% | 1 | 1 | 0% | 1,269 | 2,776 | +119% | 0 | 0 | — |
case-16 | pass→pass | 10,220 | 2,836 | -72% | 1 | 1 | 0% | 863 | 2,242 | +160% | 0 | 0 | — |
case-17 | pass→pass | 7,867 | 8,724 | +11% | 1 | 1 | 0% | 434 | 2,274 | +424% | 0 | 0 | — |
case-18 | pass→pass | 11,082 | 6,363 | -43% | 1 | 1 | 0% | 1,063 | 2,818 | +165% | 0 | 0 | — |
case-19 | pass→pass | 12,186 | 6,302 | -48% | 1 | 1 | 0% | 1,339 | 2,923 | +118% | 0 | 0 | — |
case-20 | pass→pass | 15,840 | 12,617 | -20% | 1 | 1 | 0% | 1,898 | 3,236 | +70% | 0 | 0 | — |
case-21 | pass→pass | 5,644 | 9,120 | +62% | 1 | 1 | 0% | 1,011 | 2,473 | +145% | 0 | 0 | — |
case-22 | fail→pass | 9,425 | 9,001 | -4% | 1 | 1 | 0% | 836 | 2,461 | +194% | 0 | 0 | — |
case-23 | pass→pass | 15,331 | 8,252 | -46% | 1 | 1 | 0% | 1,754 | 3,085 | +76% | 0 | 0 | — |
case-24 | pass→pass | 12,278 | 8,769 | -29% | 1 | 1 | 0% | 1,268 | 2,348 | +85% | 0 | 0 | — |
case-25 | fail→pass | 11,253 | 2,831 | -75% | 1 | 1 | 0% | 1,820 | 2,271 | +25% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted. The headline lift of +16 percentage points is the difference between those two pass rates over the 25 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.