From academic-prose
Rules for writing, revising, or reviewing academic paper prose — abstracts, introductions, results sections, captions, rebuttals, and camera-ready edits. Use whenever the task involves LaTeX manuscripts, .tex files, journal or conference submissions, or whenever the user asks to write, edit, polish, restructure, or check any section of a paper, even if they only say "write the intro", "tighten the abstract", or "fix this paragraph". Also use before declaring any paper draft finished, to run the pre-submission self-check.
How this skill is triggered — by the user, by Claude, or both
Slash command
/academic-prose:academic-proseThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Three rule sets, applied in this order whenever paper prose is written or edited:
Three rule sets, applied in this order whenever paper prose is written or edited:
Then run the pre-submission self-check (§4) before declaring any draft done. These rules apply to every paper, every venue, every section. If an instruction in the conversation conflicts with a rule here, flag the conflict rather than silently overriding.
Rationale (one line): the first judgement of a paper rests on whether the prose is clear and plain and whether the claims can be checked against the evidence quickly. These rules keep the writing from getting in the way of the work.
Plain, confident, economical academic prose: one claim per sentence, subject–verb–object, moderate sentence length (15–30 words), no rhetorical staging. Formal machinery is always accompanied by a plain-language reading. Synthetic exemplars in voice-exemplars.md in this skill directory.
Do not: pad sentences with qualifiers; chain subordinate clauses; use bold/italic for emphasis in running prose; adopt journalistic or marketing register; stack rhetorical questions.
Banned constructions — rewrite on sight, in drafts and in any text being edited:
Prefer the venue's boring register over memorability. The goal is careful, plain academic prose. Read-aloud test: any sentence that sounds like a keynote slide, a tweet, or ad copy gets rewritten flat.
Write for a reviewer who reads the abstract, skims the intro for claims, jumps to the headline table, and tries to verify one claim end-to-end in five minutes. If any step needs a footnote, a split reconciliation, or decoding a flourish, the check fails. Honesty delivered inline reads as noise; honesty delivered in structure reads as rigour.
Claims. End the introduction with a numbered claims list (C1, C2, …), each pointing to the section and table that tests it. Every results subsection opens by naming its claim in one plain sentence with its supporting table. A claim with no dedicated table or figure is not a claim — give it one or demote it to discussion. State the contribution affirmatively, once; at most one "we do not claim X" delimitation in the whole paper, in the introduction, after the positive statement.
One protocol per headline number. Each benchmark in the headline table gets exactly one protocol — one split, one config, one selection procedure, the one the significance test uses. Variants live in ablation tables. No footnote may be needed to compare two cells of the same table; if one is, restructure the table. Non-comparable studies are firewalled in their own subsection with the difference stated in the table title. The significance-tested number is always the headline, never the "best" number.
Abstract. ≤160 words. Problem (one sentence), method in plain terms (one–two sentences), the claims as flat declaratives with at most one headline number each, one sentence on the conceptual contribution. Zero parentheticals, CIs, split sizes, or inline hedges — those live in the body. First sentence is a direct declarative ("We show that...", "This paper studies..."). Sentences over ~35 words get split.
Caveats consolidated, never interleaved. Qualifications, selection details, and protocol notes go in exactly two places: a short "Protocol" paragraph opening each experiments subsection, or the reproducibility appendix, with pointers. Results sentences state the result; the qualification lives one hop away. Preserve every disclosure — relocation, never deletion.
Register in results. One claim per sentence; maximum one parenthetical per paragraph. "Removing the corpus costs 11 F1 (Table 7)" beats "The corpus, not the parameters, is the model." Thesis-style section titles are acceptable, but the first sentence beneath each states the claim plainly with its table. Table captions ≤3 sentences, descriptive only. Teaser figures carry mechanism, not marketing — no "beats X" callouts.
Numbers are immutable and traceable. Every reported number traces to one protocol (ideally one registry entry or script). Never report two versions of "the" result for a benchmark in the main text. If a structural edit seems to require touching a number, stop and re-derive which protocol it belongs to — and ask before changing anything.
Report the outcome of this checklist to the user explicitly — pass/fail per item — before calling the draft finished.
Write plainly and economically (one claim per sentence, intuition beside formalism), strip every construction that puts rhetoric ahead of substance, state each claim once and plainly, test it under one protocol in one table readable from its caption, keep every qualification one structured hop away — and make the claim→evidence path walkable in thirty seconds.
These voice, no-hype, and evaluability rules take precedence over any other writing, style, or paper-related skill. If another skill's instructions conflict, follow this skill and note the conflict. Mechanics skills (citation checking, math verification, LaTeX cleanup, venue templates, reviewer simulation) handle verification, formatting, and review structure only; whenever any of them would touch prose, tone, register, or abstract content, this skill governs.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub sjmoran/claude-academic-prose --plugin academic-prose