academic-writing-skills
Skill442 starsChanged 21 days agoMITView on GitHub
Frame and outline academic papers, plan and revise connected paragraphs, and align claims with evidence across manuscript sections, abstracts, supplements, references, and reviewer responses. Use for scoped drafting, substantive revision, proofreading, or cross-file integration. Preserve the author's research goal, terminology, clear wording, and inference limits; scale checks to the requested passage, section, or submission package.
1 other project uses a version of this, 1 of them edited.
Adopted by
Most recently taken up by
- bas1l/claude-skillsedited, first seen Nov 11, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 1
- bas1l/claude-skills★ 059% same
Wrote their own version · 0
None found.
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- academic-writing-skills-zh-cnSkill
- academic-writing-skills-zh-twSkill
- paper-reviewSkill
- paper-review-zh-cnSkill
- paper-review-zh-twSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
22 rules. About 0 are common, 21 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈21
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Treat a manuscript as an evolving evidence system, not as isolated prose.
- Distinctive: line 37
Use the ABM examples in the display guide and overlay contract as source-grounded reasoning aids, not journal-wide templates or …
- Distinctive: line 60
Read lifecycle-and-routing.md for stage gates, review maturity, non-IMRAD routing, and change-impact classes.
- Distinctive: line 73
When a substantive Results task spans displays or subsections that answer one question, use the joint-answer plan before changing …
- Distinctive: line 83
Before drafting an outline, section, or paragraph, identify its reader function, central claim or question, authorized evidence, inference …
- Distinctive: line 91
Before drafting or substantively rewriting, run the pre-draft logic check in universal-integrity.md: check both the progression within a …
- Distinctive: line 109
Always consider these gates within the requested scope and directly affected dependencies.
- Very specific: line 118
Release integrity: never use FINAL, VERIFIED, or SUBMISSION_READY while a high-severity blocker or unresolved required source remains.
- Distinctive: line 121
Read universal-integrity.md for paragraph, abstract, conclusion, evidence, terminology, citation, and four-pass review rules.
- Distinctive: line 142
Keep overlay rules classified as:
- Distinctive: line 150
Never promote an overlay rule into a universal rule.
- Distinctive: line 156
Verify that the named overlay is available before claiming it ran.
- Distinctive: line 165
Use its comment ledger, response-order, evidence, cross-reference, and formatted document checks together with the active manuscript state …
- Distinctive: line 184
Do not make a global replacement across equations, field codes, citations, XML, or metadata without protected-context checks and a …
- Distinctive: line 198
Treat script findings as diagnostics.
- Distinctive: line 205
check added, removed, or semantically affected citations against their claims and the reference list.
- Distinctive: line 206
run terminology, forbidden-variant, repetition, stock-phrase, and project-discouraged-phrase checks on the exact candidate itself.
- Distinctive: line 229
If a later pass finds a material issue, fix it and repeat every affected upstream and downstream check.
- Distinctive: line 233
After every review, revision round, audit, or release check, explicitly revisit the task's functional coverage before reporting completion.
- Distinctive: line 247
Read reviewer-red-team-and-release.md for negative-space tests, issue severity, readiness criteria, and the final report format.
- Distinctive: line 253
Use the dedicated document, spreadsheet, presentation, or PDF skill for file-format operations and visual verification.
- Distinctive: line 265
Do not call a manuscript final when the evidence supports only a partial or conditional review.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Academic Writing Skills
Governing Principle
Preserve the authors' evidence, intended meaning, and locked decisions while making the scholarly argument clear, reproducible, internally consistent, and no stronger than the evidence permits. Never invent results, mechanisms, citations, metadata, analyses, or field requirements.
Treat a manuscript as an evolving evidence system, not as isolated prose. Follow the author's writing sequence, but re-open earlier sections whenever later evidence changes them.
Route the Task
First identify:
- requested action: plan, draft, review, revise, proofread, or certify readiness
- artifact scope: sentence, paragraph, subsection, chapter, full manuscript, supplement, or submission package
- change impact and available verification sources, separately from artifact scope
- manuscript archetype and study design
- lifecycle stage and review maturity
- authoritative sources, active files, and historical references
- user-locked wording, facts, terminology, and decisions
- whether editing is authorized or the task is diagnosis only
Use lightweight mode for bounded text work. Use managed-project mode for full-manuscript integration, companion-file coordination, or release certification; reuse an existing managed context for local work in that project. Managed context does not turn a paragraph edit into a whole-paper audit. Route checks by requested action, textual scope, change impact, and source availability using lifecycle-and-routing.md.
In managed-project mode, read state-and-authority.md and reuse equivalent project records. Initialize the project-state template with scripts/init_manuscript_state.py only when sustained coordination actually needs a record and none exists. Do not create project machinery for a one-off edit or reinitialize state merely because a new turn begins.
When project terminology is supplied or sustained writing needs a shared vocabulary, use the terminology procedure in state-and-authority.md, including for local work in that project. Read the relevant confirmed entries before writing; keep new suggestions separate from confirmed choices and distinguish configured scan results from vocabulary completeness.
Confirm Venue Direction, Format, and Style
For venue-directed planning, substantive restructuring, display/reference formatting, or submission checks, first confirm the target journal, article type, audience, and publication stage. Use the Venue Profile in overlay-contract.md to separate scientific direction, verified requirements, stylistic examples, and project preferences. Reuse a current profile; verify changed or uncertain official requirements before applying them. A grammar-only edit does not trigger journal reselection or a whole-package audit. If no venue is selected, keep advice provisional and do not invent compliance.
Use the ABM examples in the display guide and overlay contract as source-grounded reasoning aids, not journal-wide templates or certification of the original paper.
For an extended-outline request, read the paragraph-opening format and worked skeleton in professor-teaching-cases.md. Follow an author-supplied example first; make the argument readable through actual or proposed paragraph-opening sentences, with citations and evidence needs attached where relevant. Use the other cases there for a requested abstract, reviewer-response, or display demonstration; do not load all examples or expand a grammar-only edit into a section review.
Follow the Manuscript Lifecycle
Map the work to this lifecycle without forcing the final article order to equal the writing order:
- establish project authority, manuscript archetype, and intended venue
- develop the extended outline, Introduction, gap, questions, and planned contribution
- develop study context, data, theory, and Methods
- build Results from verified analyses, figures, tables, or source evidence
- develop Discussion, alternatives, implications, limitations, and future work
- synthesize the Conclusion from the stabilized evidence map
- rebuild the title, Abstract, highlights, and summaries from the current manuscript
- reconcile references, parallel-developed supplements, metadata, and submission files
- pass release checks on the exact deliverables
At every stage, test the relevant part of both directions of alignment:
- top-down: purpose or gap → question or objective → method → expected evidence
- bottom-up: source evidence → result → interpretation → contribution and summary claim
Read lifecycle-and-routing.md for stage gates, review maturity, non-IMRAD routing, and change-impact classes.
For planning, drafting, or reorganizing an Introduction or study-context section, read introduction-and-study-context.md. It guides section functions and content placement, not fixed paragraph counts or a mandatory article structure. A bounded language edit uses the local prose checks instead.
For planning or developing Methods, or substantively clarifying a setting's source, rationale, or assumption status, read methods-and-assumptions.md. Distinguish the provenance of a value or rule from assumptions about its application and from evidence of validation. Use the hypothetical examples as reasoning aids, not as verified study facts.
Load the remaining guides only for the requested reader function or an affected dependency:
- results-and-explanation.md: evidence-led findings, so-what/SQ closure, visual functions, and supported versus merely programmed mechanisms.
- discussion-and-limitations.md: interpretation or extension-analysis mode, sensitivity boundaries, and specific-constraint-first limitations.
- conclusion-and-abstract.md: recover the current research goal, work, comparison levels and scope before summary drafting; preserve clear author wording and distinguish study identity from findings.
- figures-tables-and-supplements.md: final-size legibility, captions/notes, parallel SM, numbering, and project formatting.
When a substantive Results task spans displays or subsections that answer one question, use the joint-answer plan before changing individual paragraphs. A bounded selection may need those sources as reading context; it does not authorize wider edits.
For a proposed interpretation from an advisor, reviewer or external AI, including in a rebuttal, apply the current-evidence checks in results-and-explanation.md.
At each chapter completion or substantial chapter revision, run the chapter-closure blind-spot check in lifecycle-and-routing.md. The checkpoint is required; select diagnostic questions by design, content, and evidence. Do not run a full chapter audit for a grammar-only edit or claim exhaustive discovery of unknown unknowns.
Draft from Explicit Writing Contracts
Before drafting an outline, section, or paragraph, identify its reader function, central claim or question, authorized evidence, inference limit, and link to adjacent material. For a paragraph, use this compact contract:
- function: what the paragraph must accomplish
- claim: the narrowest defensible main point
- evidence: the source, result, citation, or reasoning that supports it
- development: how the evidence is explained without adding a mechanism
- bridge: what relation leads into the next paragraph
Before drafting or substantively rewriting, run the pre-draft logic check in universal-integrity.md: check both the progression within a paragraph and the relation between adjacent paragraphs. Match the depth to the requested scope; a local edit does not require replanning the whole section.
Draft only after the contract and information order are coherent. After drafting, compare the paragraph against the contract and re-read its previous and next paragraphs. Do not polish a paragraph into fluency if its function, evidence, or placement is wrong.
For comment- or selection-driven edits, first apply comment-scoped-revision.md: verify the original/current selected text and full thread, separate reading from editing scope, and compare the exact candidate with the authorized boundary. Assess whether the original concern is answered separately from whether wording changed. Do not substitute the whole paragraph for the selection or treat a broader explicit instruction as title-only because its anchor is a heading.
For substantive drafting or formatted-file editing, read execution-and-acceptance.md before drafting. Use its reader-route and caption/body/SM allocation, then perform its scoped checks before the first handoff. Preserve the author's primary comparison before dependent subgroup or model detail; this does not prescribe a universal comparison order.
Use that guide's stage checklist to connect outline, pre-draft planning, post-edit checks, chapter closure, and full review to concrete records. Keep local handoffs concise and expand the record at chapter/full-review boundaries; reuse the shared project state rather than creating another glossary or checklist system.
Make sentence-to-sentence reader continuity a high-priority acceptance gate before the first handoff and after any substantive revision. For process descriptions, use the workflow handoff test; keep reader findings distinct from scan results. For limitation prose, also apply the constraint-versus-defect triage in discussion-and-limitations.md: describe a documented boundary and its consequence; correct fixable errors instead of recasting them as limitations. Keep this expanded gate out of a grammar-only local edit.
Also identify the intended reader's likely knowledge. For interdisciplinary papers and reviewer responses, assume a reader who understands the broader research area but does not know the study's internal terminology, model sequence, data transformations, or earlier drafting history. Supply the minimum context that reader needs at the point of use; do not make the reader reconstruct it from later text.
For an explicit independent readability review, a concept-dependent substantive section, or repeated comprehension failures, load independent-reader-review.md. Use its scoped cold-reader procedure and worked examples to catch unexplained terms at first use; do not impose multi-agent review on a grammar-only edit or present AI reader roles as human validation.
Apply Universal Integrity Gates
Always consider these gates within the requested scope and directly affected dependencies. Determine applicability; do not rerun every whole-paper check for every local edit:
- Authority: derive facts from identified sources; do not infer authorship roles, funding, sample sizes, or methods from formatting or old drafts.
- Contract: preserve locked gap, task, questions, outcomes, contribution, and explicit nonclaims unless the user authorizes a semantic change.
- Alignment: connect the affected question or objective to its method, evidence, substantive answer, and inference boundary; check the complete map during whole-paper review.
- Method–evidence integrity: require enough detail to understand how evidence was produced and which assumptions bound the inference.
- Claim scope: distinguish direction, magnitude, uncertainty, significance, equivalence, causation, mechanism, prediction, and generalizability.
- Cross-artifact synchronization: propagate material changes to every affected section and companion artifact.
- Version integrity: identify one active release and distinguish it from historical references.
- Release integrity: never use
FINAL,VERIFIED, orSUBMISSION_READYwhile a high-severity blocker or unresolved required source remains. - Prose integrity: preserve one term per concept, check cross-disciplinary comprehension and noun-chain clarity using prose-and-citation-editing.md, distinguish necessary technical repetition from avoidable verbal repetition, and verify paragraph-to-paragraph flow without using synonym rotation to hide repetition.
Read universal-integrity.md for paragraph, abstract, conclusion, evidence, terminology, citation, and four-pass review rules.
Select Study-Design Adapters
Select adapters by research design, not by discipline label. Load only the relevant sections of study-design-adapters.md:
- quantitative observational or survey research
- experiments or quasi-experiments
- qualitative research
- mixed methods
- computational models or simulations
- AI/LLM-based studies
- evidence syntheses, reviews, or meta-analyses
- theoretical, conceptual, framework, methods, or data papers
Combine adapters when the design is genuinely hybrid. Treat adapter checks as questions requiring manuscript evidence, not assumptions that a method was used incorrectly.
Compose External Overlays Safely
Apply a supplied domain skill, reviewer rubric, professor checklist, reporting standard, or journal guide as an overlay. Read overlay-contract.md before applying one.
Keep overlay rules classified as:
- study-design requirement
- domain convention or technical check
- reviewer preference or tone
- venue requirement
- project-specific fact or decision
Never promote an overlay rule into a universal rule. When an overlay conflicts with evidence, author instructions, ethics, or defensible inference, preserve the evidence and report the conflict. A request to explain “why” does not authorize an unsupported mechanism.
Apply Registered Project Workflows
When the project, folder, manuscript, or user request identifies Survey paper, PiDS, or a WC_vX manuscript, use the installed survey-paper-wcvx skill as the project-state and reviewer-round overlay together with these universal integrity gates. Keep that project's filenames, author identity, priority scheme, and version rules within the project overlay.
Verify that the named overlay is available before claiming it ran. If absent, use supplied authoritative project records for the scoped task and disclose the missing overlay; do not invent its rules or attribute earlier failures to it.
Handle Reviewer Responses
For rebuttal letters, response-to-reviewers documents, or revisions of an advisor-edited response draft, read reviewer-response-workflow.md. Use its comment ledger, response-order, evidence, cross-reference, and formatted document checks together with the active manuscript state and any supplied lab examples. Keep reviewer-facing prose, manuscript or supplement changes, and internal replies to an advisor distinct. A high-level synthesis response may navigate to later detail, but the first substantive response to an issue must explain the change, evidence, interpretation, manuscript effect, and remaining boundary needed to answer that comment.
Control Material Changes
Classify each proposed change before editing:
- Class A — semantic contract: gap, task, question, outcome, contribution, causal or validation framing
- Class B — evidence or method: data, sample, model, analysis, number, figure, table, limitation
- Class C — metadata: authors, affiliations, funding, roles, declarations, repository links
- Class D — surface: grammar, punctuation, formatting, local clarity with no scientific change
Run the corresponding impact scan before declaring the edit complete. Class A normally reopens the whole evidence chain and all summaries. Class B reopens affected methods, results, interpretations, visuals, supplements, and summaries. Class C reopens every submission artifact and metadata field. Class D requires a local semantic diff and formatted-file check.
Do not make a global replacement across equations, field codes, citations, XML, or metadata without protected-context checks and a post-change diff.
Use Deterministic Audits as Evidence, Not Judgment
For managed projects, use the bundled scripts when relevant:
scripts/audit_manuscript_state.py: validate state completeness, source links, question alignment, dimension coverage, and release blockersscripts/audit_text_consistency.py: extract text from supported files and scan registered locked strings, prohibited variants, and fact conflictsscripts/audit_prose_patterns.py: report exact duplication, repeated openings and phrases, stock phrasing, and candidate nontechnical word overuse without claiming AI authorshipscripts/audit_candidate_text.py: scan the exact proposed passage against the active project terminology, forbidden variants, and prose profile and record its hashscripts/audit_handoff.py: optionally validate scope-dependent coverage records and exact-file hashes;COVERAGE_COMPLETEis not editorial or scientific certificationscripts/audit_docx_structure.py: inspect Word OOXML using exact tag names for tracked changes, revision authors, comments, parent-linked replies, reply authors, fields, and placeholdersscripts/run_regression_tests.py: verify the audit tools against bundled failure cases
Treat script findings as diagnostics. Inspect each match in context before editing. A clean script report never replaces substantive reading.
Gate the Exact Candidate Before Delivery
After the last wording change, freeze the exact passage or artifact that will be delivered. Reuse the relevant active-state entries and applicable overlay rules, including for local work in a managed project. Do not load unrelated project history or audit companion files merely because they exist. Then, at the depth required by the requested action and change impact:
- compare the exact candidate with its writing contract, locked meaning, evidence, and adjacent paragraphs;
- check added, removed, or semantically affected citations against their claims and the reference list; substantive passage review also checks retained claim–citation support, while grammar-only work preserves unchanged verified support and states any verification limit;
- run terminology, forbidden-variant, repetition, stock-phrase, and project-discouraged-phrase checks on the exact candidate itself;
- inspect and resolve each deterministic finding in context, or record an explicit author-approved reason to retain it; and
- record which candidate was checked, using its hash when a deterministic candidate audit is available.
Before the first draft handoff, record concrete, located findings for reader comprehension, cumulative progression, and every affected caption, saved layout, venue-format basis, and new edit/reply identity using execution-and-acceptance.md. Keep lightweight records inline or reuse existing project records; JSON is optional. Missing or unavailable checks remain incomplete, not passed. A language-scan PASS never substitutes for these checks. Provisional drafts may be delivered with explicit limits, but not as fully checked or journal-compliant artifacts.
For prose revision, select the applicable functional checks in prose-and-citation-editing.md: transitions, function-preserving concision, natural scholarly prose, claim–citation alignment, observable stock or AI-like features, and dash/hyphen style. A punctuation edit needs meaning preservation, not a new literature audit. Treat diagnostics as prompts for judgment, not universal bans or authorship detection.
For a reference audit or changes to numbered citations, use the two-way reconciliation in prose-and-citation-editing.md, even when no prose revision is requested.
For a substantial manuscript section, abstract, rebuttal, or explicit banned-word audit, also inspect the exact candidate against banned_words.md. Treat every match as a contextual diagnostic rather than an automatic deletion or evidence of AI authorship, and apply any project-specific writing contract or terminology override first.
Any edit after this gate invalidates its result. Re-run every affected candidate check before calling the wording checked, final, polished, or ready to paste. Never report that academic-writing-skills passed when only an earlier draft, the surrounding manuscript, or an unchanged source artifact was audited.
Conduct Four Distinct Passes
For whole-paper or submission-package review, complete four top-to-bottom passes. For a substantial subsection or chapter revision, apply the passes to that unit and its affected dependencies; a bounded language edit uses the local candidate gate:
- Argument and structure: reader functions, purpose, questions, organization, paragraph openings, and contribution.
- Evidence and scope: methods, data, results, figures, tables, citations, uncertainty, claim strength, and cross-file propagation.
- Scholarly writing: clarity, natural subjects, terminology, necessary versus avoidable repetition, sentence-opening variety, observable stock phrasing, paragraph flow, syntax, tense, voice, and notation. Never label prose AI-generated from style alone.
- Delivery integrity: summaries, references, numbering, metadata, exact filenames, tracked changes, visual rendering, and release blockers. Read the revised prose aloud or simulate an ordinary spoken reading. If a sentence would sound unnatural when said by a researcher, or its meaning becomes clear only after rereading, revise its syntax, referents, or information order without making the scholarly language casual.
If a later pass finds a material issue, fix it and repeat every affected upstream and downstream check. Do not certify the earlier pass as current.
End Every Check with a Functional-Completeness Retrospective
After every review, revision round, audit, or release check, explicitly revisit the task's functional coverage before reporting completion. Ask:
- Did the work cover the requested lifecycle stage and every artifact in scope?
- Did it preserve locked meaning and consult the correct authority sources?
- Did every affected question, method, result, interpretation, limitation, and summary remain aligned?
- Did all relevant study-design adapters and overlays run without becoming universal assumptions?
- Did the change-impact scan cover upstream, downstream, and companion-file dependencies?
- Are any unknowns, unavailable sources, open issues, waivers, or high-severity blockers still present?
- Were deterministic, structural, and visual checks used where applicable, and were their limits stated?
- Is the exact deliverable—not an intermediate copy—the one inspected?
- Did the exact post-edit candidate pass the applicable project profile, and were all findings resolved or explicitly retained with a reason?
Apply these questions to the agreed scope, not every possible artifact. If a required answer is no or unknown, continue safe in-scope work or report the limit. Distinguish checked, not applicable, outside requested scope, and source unavailable/unverified; an excluded check is not a pass. Never collapse this retrospective into a generic “all checks passed.”
Read reviewer-red-team-and-release.md for negative-space tests, issue severity, readiness criteria, and the final report format.
Work with Formatted Files
For a Word handoff or layout review that covers data tables, use the per-table print-layout check in figures-tables-and-supplements.md and record it in the formatted-file acceptance check. A narrow wording edit with no table or pagination impact does not trigger a package-wide table audit.
Use the dedicated document, spreadsheet, presentation, or PDF skill for file-format operations and visual verification. This skill governs scholarly integrity; it does not replace format-specific rendering, tracked-change, or OOXML workflows.
Report the Outcome
Lead with readiness and remaining blockers. Report only checks actually performed. State:
- what changed and what was preserved
- whether scientific meaning, evidence, numbers, citations, or metadata changed
- which artifacts and lifecycle gates were checked, not applicable, outside scope, or unverified
- unresolved issues, unavailable sources, or explicit waivers
- the functional-completeness retrospective outcome
Do not call a manuscript final when the evidence supports only a partial or conditional review.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

