CommunityEscrita e Ediçãogithub.com

academic-writing

Revise academic prose for clarity, as a focused edit or an optional scored rewrite loop with local stylometry. Use for manuscript prose, abstracts, grants, response letters, scientific reports and research pages.

O que é academic-writing?

academic-writing is a Claude Code agent skill that revise academic prose for clarity, as a focused edit or an optional scored rewrite loop with local stylometry. Use for manuscript prose, abstracts, grants, response letters, scientific reports and research pages.

Funciona com✓Claude Code~Codex CLI~Cursor✓Gemini CLI
npx skills add https://github.com/fmschulz/academic-writing-skill/tree/HEAD

Installed? Explore more Escrita e Edição skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Perguntar na sua IA favorita

Abre um novo chat com esta habilidade de agente já pré-carregada.

Documentação

Academic Writing

Revise stiff, generic, unclear or AI-shaped prose into clear academic writing without locking it to a field. The skill uses clarity checks, source-derived sentence frames, local stylometry and meaning checks. A frame supplies fixed syntax, never the source article's facts or an author's voice. stop-slop owns sentence clarity and AI patterns for all prose; for scientific text, apply it together with this skill. Use a lighter edit for casual writing, code comments, changelogs and mechanical formatting.

Priorities

Resolve every conflict in this order: meaning (facts, numbers, citations and the strength of each claim stay as the source has them), clarity (a reader in the field understands each sentence on one reading), genre (the section follows its conventions and the venue's rules), naturalness (rhythm, variety and the local scores). A lower priority never overrides a higher one. Do not trade a clear sentence for a better score, and do not change a claim to make a sentence clearer.

Focused edit

This is the default. Audit the baseline, review the candidate, then resolve findings and review any revision. Without a coverage request, a clean draft needs no further edits. The scaffold pass requires independent review, including in this default mode. Other focused edits may use identified self-review unless the task requires more. Apply the stop-slop cold-read conditions for standalone deliverables when the harness can start a fresh subagent. Do not require a profile, style card, scorecard, six review lanes or a whole-manuscript rewrite.

  1. Read the passage and enough surrounding text to know its job, its audience and the terms the document already defines. For project files, technical checking or source integration, read manuscript safeguards.
  2. Lock the facts: every number, unit, threshold, citation, equation, quoted term, source label, named entity and explicit limitation. This list holds for every edit, scored or not. Before editing, record located findings and evidence gaps against the original source in task notes, using the claim checks.
  3. Run the three checks in clarity.md on each sentence, with its scientific prose section. The studied object, its measure or we is the subject and the finding is the main verb; the studied object is never the agent of an observed difference. Every term, reference and number resolves for a reader in the field, with no working label, noun stack, bare This or comparison without a baseline. The result is in the main clause. Take a missing actor, definition or quantity from the manuscript and its evidence; never from a donor article. Identify the relationships the sentence needs: ranked objects and ranking criterion; compared objects, measure and baseline; or the object undergoing a method and the operation performed. Make a builder or other actor explicit when the manuscript names it and its relation to those objects is unclear. Use the surrounding text where it already resolves a relation; do not make every sentence repeat the context. Record located findings, not an empty checklist for every sentence. A longer sentence is preferable when it states these relationships clearly. Preserve necessary technical repetition and effective sentences. Every added actor, criterion, measure, baseline or other relation needs support in the manuscript or its evidence. If that support is missing, retain the affected claim and report the gap; never guess. Meet a supplied word limit by cutting dispensable material, following clarity over brevity, without making a necessary relationship implicit.
  4. Review the cues in AI tells, including its scholarly exposition patterns and contribution, citation, aggregation and proposal checks.
  5. Check the result against the locked facts and the strength of each claim. Flag visible technical errors.
  6. For sentences the AI creates or rewrites at the clause level (predicate, clause order or sentence structure), edits that add or change a scientific relationship even within a phrase, or every paragraph in an explicit coverage request regardless of authorship, run the required scaffold pass. Select compatible frames from data/scaffolds.jsonl or a verified task-local bank with ranked list --query searches for the intended predicate, hedge or relation. Search across families before narrowing; try supported alternative predicates when the first search is unhelpful. Let identified sentence relationships guide retrieval; topical similarity alone is insufficient. Fill only their named slots, and render them with scripts/scaffolds.py. Use alternatives with the same rhetorical function; use a frame when it improves readability or precision, or supplies requested structural variety while preserving meaning and clear paragraph flow. Without a requested coverage target, the plain draft is the default and zero frames is valid. A requested number per paragraph instead uses the coverage contract; a few examples do not complete a manuscript-wide request. Keep fixed frame spans locked during later edits. Missing tools or data make this pass incomplete, not a synthetic fallback.
  7. Review the exact final text for meaning, readability, flow and style, using the independent scaffold review when step 6 applies. Resolve findings, revise where useful and review again after the last change. Try up to three candidates per passage by default; then retain the original baseline or clearest reviewed candidate and review it too. An explicit coverage target requires continued retrieval under the coverage contract. Never substitute synthetic profile or example content. Always disclose unresolved source gaps. The scored loop uses its own four-round total budget, not a nested loop. When step 6 applies, run the final file check last, after review, with the required review record. Zero replacements completes only a pass without a coverage target; a target also requires the coverage table with qualifying structural changes judged against the original manuscript. Track clarity findings separately from scaffold counts: meeting a paragraph's replacement target never closes an unresolved clarity finding. Resolve every finding or disclose the source gap that prevents a supported repair.

Keep the purpose, audience, voice, structure, document class, bibliography and notation unless the task changes them. Invent no length limits, sections or extra deliverables. Methods, legal, statistical and procedure sections follow their genre first: allow natural passive voice and keep exact terms, and limit the clarity pass to undefined terms, unclear referents and actions hidden in nouns. Stop when the requested passages are clear and checked, with unresolved gaps named. If a coverage target remains unmet, report the pass as incomplete even when the prose and exact rendering checks pass. Coverage never overrides meaning or clarity, and a coverage request alone does not start the scored loop.

Scientific research pages

Use this workflow for scientific reports, results summaries and project research pages. It takes precedence over the scored loop and the output contract. A routine operational status update needs a proportionate edit, not the source-frame review cycle. A scientific correction that changes a relationship follows the focused edit and its review requirements.

Default to a reviewable static page in the project's existing site, in its source format, at an approved destination and access model. An explicit format or hosting choice wins; this default authorizes no deployment.

  1. Read this skill and stop-slop before drafting. Apply stop-slop to the source, then the focused edit, then stop-slop again on the finished prose. Keep selected frame text fixed through this last pass and review and validate the final artifact after it.
  2. Write concise prose. State the result or question directly, name the subject in headings, and order methods, results, interpretation and proposed work. Use only the sections the content needs. Cut sales language, slogan headings, decorative number displays, repeated caveats and commentary on the writing.
  3. Keep limitations beside the results they qualify, and completed analyses apart from proposals. Cut repetition without turning missing annotations into missing genes, association into mechanism, or planned work into results.
  4. Name both skills and the concise-scientific-prose requirement in every delegated drafting or rewriting task, for example: "Read and apply stop-slop and academic-writing. Write concise, well-structured scientific prose; preserve facts and uncertainty." Check returned prose against the evidence. The lead owns the final edit.
  5. New or substantially rewritten results and interpretation need an independent editorial review by a fresh reviewer, ideally from another model family, with a prose review prompt. Give it the text and this rubric: quote each sentence you read twice; label the fault (hidden actor or action, vague wording, buried point); give a rewrite that keeps the claim; flag any change in meaning. Verify factual suggestions before applying them. Read the final rendered page.
  6. Before first publication or a substantive update, record use of both skills, the review, resolved findings and the source and render checks in the project's methods or task record. Report an incomplete review or check as incomplete. This gate grants no permission to publish.

A phrase scan or style score cannot show that prose is clear, accurate or natural; scoring is optional here. Deliver the page with a short completion note. Keep scorecards, agent discussion and review records in tasks/.

Scored loop

Use council-loop.md only for an explicit scored revision, or when a substantial revision benefits from it. It adds a style profile, an optional style card, six review lanes, a rewrite brief and a scorecard. The scores are revision aids: the draft's purpose and claims outrank them, a passing score is not editorial approval, and no score measures clarity (the reader-clarity lane does).

Writing review

For a writing review of a finished manuscript (review modes, figures and legends, term and number consistency, citation tracing), read manuscript-review.md. It does not judge the science; use nelli-paper-review for peer review.

Boundaries

  • The task supplies the content. Import no facts, examples or claims from a source article. Use its accepted sentence framework only when every fixed predicate, hedge and relation fits the draft's supported claim. Field-specific verbs may remain fixed in a suitable frame; they must not add a method or mechanism the task does not support. Optional teaching shapes live in references/academic-style/.
  • Integrity comes first. If a sentence cannot improve without changing the claim, keep the original clause and report the residual issue. A requested substantive correction is separate from polishing: verify it, and make it explicit.
  • Protect effective prose. Keep one term per concept, short sentences that state findings, and the links between sentences. These outrank a style card or vocabulary metric.
  • Plain prose is never a defect. Reading grade, passive ratio, transition density and hedge density are ceilings only.
  • Naturalness is a band, and scores are proxies. A signal is a human envelope, never a quantity to push to an edge: predictable text and text randomized past sense both read as unnatural. Use real writing quality only, never invisible Unicode, homoglyphs or injected typos. Do not claim a draft will pass GPTZero, Turnitin or any detector, or that output evades detection.
  • Source frames require exact rendering. Use only accepted bank entries and keep every fixed character. Change only declared slots. Synthetic teaching examples and profile-corpus exemplars are not approved frames. The validator proves fixed-text fidelity for checked files, not meaning, grammar, authorship or that an agent followed the skill. A style card never overrides the writer's voice and logic (scaffold pass).
  • The author owns the claims. If any input was AI-drafted, the author follows the venue's disclosure policy.

Resources

PathRead when
references/ai-tells.md, stop-slop's references/clarity.mdSteps 3 and 4 of the focused edit.
references/council-loop.mdManuscript safeguards; the scored loop, its stop criteria and output.
references/skeleton-injection.mdRequired source-frame pass for AI-created or clause-level rewritten sentences, changed scientific relationships, and all paragraphs in an explicit coverage request.
references/reviewer-roles.md, rewrite-backends.md, templates/Scored loop: lanes, optional rewrite backends (Gemini via CBORG, Claude via claude -p), brief, scorecard.
references/academic-style/Scored loop or a new draft: metrics_targets.md, style-guide.md, checklist.md.
scripts/stylometry.pyscore, tells, sample (style card), selftest.
scripts/profile_corpus.pyBuild optional style profiles and shape examples: --fallback, --flavor, --corpus, --roles. These do not satisfy the source-frame pass.
data/scaffolds.jsonl, scripts/scaffolds.pyRequired source-frame selection, rendering, final-file and review-record validation, and generated MCP export.
scripts/build_scaffolds.pyExtract review candidates from verified original article text; never accepts frames automatically.
scripts/cborg_rewrite.py, claude_rewrite.shOptional candidate generators; re-score and fact-check their output.

Output contract

For a focused edit, return the requested passage or edited files and a brief note only for unresolved material issues or substantive corrections. Add no scorecards, extra sections or routine editing report. Keep the source format. For a scored revision, follow the output contract of the loop.

Habilidades Relacionadas