--- name: unslop-academic-writing description: Remove AI slop from research writing so papers, theses, grant proposals, reviews and rebuttals read as written by a careful human expert. Covers stock vocabulary (delve, tapestry, pivotal, underscores), empty emphasis, hedge stacks, formulaic signposting, "not only X but also Y" constructions, em-dash pileups, monotone rhythm, and claims vaguer than the data. Use when drafting or revising academic text with an AI assistant, when a draft "sounds like ChatGPT", before submission, or to match an author's own voice. Includes a zero-dependency slop linter for Markdown, LaTeX, text and Word files. license: MIT compatibility: Python 3.9+ standard library for scripts/slop_check.py. metadata: version: "1.0" category: research-writing maintainer: Kalaris Labs author: Kalaris Labs (Sayan Chowdhury) tags: AI slop, academic writing, tone, voice, humanize, clarity, editing, scientific writing --- # Unslop Academic Writing Slop is prose that is fluent but says less than it seems to. Readers, reviewers and editors now recognize it on sight, and it costs credibility even when the science is sound. The fix is **specificity**: every sentence should carry information a generic paper on the topic would not. This skill is about **clarity and authorship, not about disguising AI use**. Follow the venue's policy on AI assistance (most require disclosure, see `reproducibility-statement`). The author remains responsible for every claim. ## Workflow 0. **Ask what each paragraph claims.** Before any rewording, list the paragraphs that make no checkable claim: ```bash python scripts/slop_check.py manuscript.md --claims ``` It flags paragraphs of 40+ words with no number, citation, figure or table reference, acronym or named method. For each one, answer "what is the claim, and what is the evidence?" in one sentence. If you can, rewrite the paragraph around that sentence. If it only restates the field, cut it or merge it into a paragraph that has a point. If the claim exists but the author has not supplied the evidence, ask for the number or citation and leave a placeholder. Cleaning the phrasing of a paragraph with no claim just hides the problem. The flag is a heuristic: some sound paragraphs (a definition, a transition) are legitimate, so judge each one. 1. **Measure.** Run the linter on the draft (or a section): ```bash python scripts/slop_check.py manuscript.tex python scripts/slop_check.py intro.md --json > slop.json ``` It ignores math, citations, code and markup. Index bands: < 5 clean, 5-15 some slop, 15-30 heavy, > 30 severe. Each finding has a line number and a rewrite hint. 2. **Fix content first, words last.** Work top-down. Polishing vocabulary in a paragraph that should not exist is wasted effort. | Level | Question | Typical fix | |---|---|---| | Claim | Does each paragraph make one claim a specialist would find non-obvious? | Cut paragraphs that restate the field; merge duplicates | | Evidence | Is every claim tied to a number, citation, figure or method? | Replace "significantly improves" with the effect size, interval and test | | Structure | Does the logic carry the connection, or do signposts ("Moreover") fake it? | Reorder sentences so each follows from the last; delete connective filler | | Rhythm | Do sentence lengths vary with content? | Short sentence for the key finding; longer ones for mechanism and caveats | | Words | Stock vocabulary, hype, stacked hedges? | Use the hints and [references/patterns.md](references/patterns.md) | 3. **Preserve what matters.** Never change numbers, units, citations, technical terms, defined abbreviations or the strength of a claim while "unslopping". Hedges that reflect genuine uncertainty stay. Remove only the *second* hedge in a stack. 4. **Match the author's voice.** Ask for 1-3 pages the author wrote before using AI (a previous paper or thesis chapter). Note their sentence length, use of "we"/"I", preferred transitions and terminology, then keep the revision within that range. Follow the discipline norms in [references/discipline-voice.md](references/discipline-voice.md). 5. **Re-measure and read aloud.** Target an index under 5 for abstracts and introductions, under 10 for methods-heavy sections. The linter is a floor, not a judge: finish by reading the text as a skeptical reviewer. ## Rewrite rules (apply in order) 1. **Delete the frame, keep the point.** "It is important to note that X" → "X". 2. **Replace abstraction with the specific thing.** "a wide range of tasks" → "12 of the 14 benchmark tasks". 3. **Replace evaluation with evidence.** "a remarkable improvement" → "a 9.4-point improvement (95% CI 7.1-11.7)". 4. **Name the mechanism, not the role.** "plays a pivotal role in" → "is required for", "accounts for 40% of". 5. **One hedge, sized to the evidence.** "may potentially suggest" → "suggests" (strong data) or "is consistent with" (weak data). 6. **State the contrast directly.** "not only X, but also Y" → "X and Y", or keep only the stronger point. 7. **Use the verb.** "perform an analysis of" → "analyse"; "provide a description" → "describe". 8. **Let structure replace signposts.** Delete "Moreover/Furthermore/Additionally" where the logical relation is clear. 9. **Punctuate normally.** Most em dashes become commas, colons, parentheses or full stops. 10. **Vary rhythm deliberately.** After two long sentences, write a short one that states the takeaway. ## Before and after > **Before:** In today's rapidly evolving landscape, single-cell technologies play a pivotal role in > unraveling the intricate tapestry of cellular heterogeneity. It is important to note that our novel > framework significantly outperforms existing methods, paving the way for groundbreaking discoveries. > > **After:** Single-cell RNA-seq atlases now exceed 10 million cells, but cluster labels still disagree > between annotation tools on about 20% of cells. Our method reduced this disagreement to 7% on three > held-out atlases (paired t-test, p < 0.01), mainly by resolving T-cell subtypes that marker-based tools merge. ## Integrity rules - Never fabricate specifics to replace vague text. If the draft lacks the number or citation, ask the author or mark a placeholder such as `[N = ?]`. - Verify that each rewritten sentence says the same thing as the original, at the same strength. ## What the linter cannot judge The `--claims` pass finds paragraphs with *no* visible evidence; it cannot tell whether a paragraph that has a number or a citation is saying anything worth saying. Accuracy, novelty, whether a citation supports the sentence, or whether the argument is right. Pair this skill with `citation-verification`, `scientific-critical-thinking` and a human co-author's read. ## Related skills `scientific-writing`, `ml-paper-writing`, `abstract-and-title`, `rebuttal-and-response-to-reviewers`, `cover-letter-to-editor`, `peer-review`, `reproducibility-statement`.