agentleFS
unravel11unravel11/research-paper-skills/.claude/skills/paper-style-profiler/SKILL.md

paper-style-profiler

Skill0 starsChanged 27 days agoMITView on GitHub

Discover the strongest published style-anchor papers already present in the current academic repository, select a complementary three-to-five-paper corpus, extract and save its aggregate language style, refresh stale profiles, audit a manuscript against the saved contract, or apply that contract while preserving claims and terminology. Use for repository-specific paper writing and revision; not for literature review, copying an author's sentences, authorship-detector evasion, or unsupported claim strengthening. 中文触发短语:从文献库里挑几篇优秀论文提炼文风或风格合同、写作合同,按已保存的风格改写这一节但不改数字和结论,刷新文风画像,审一下这稿和范文的语言差距。

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 8 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Treat measurements as descriptive evidence, never quotas, acceptance predictors, quality scores, AI scores, or authorship classifications.

    line 15
  2. Distinctive:

    Do not optimize for detector evasion.

    line 16
  3. Distinctive:

    Use repository papers by default.

    line 17
  4. Very specific:

    If the request is ambiguous, prefer AUDIT over mutation.

    line 29
  5. Distinctive:

    Follow repository-specific source-of-truth and commit rules.

    line 41
  6. Distinctive:

    Keep raw PDF text in an explicit temporary directory outside the repository.

    line 43
  7. Very specific:

    If authority or the target manuscript is materially ambiguous, stop with BLOCKED_AUTHORITY_AMBIGUOUS and name the missing decision.

    line 45
  8. Very specific:

    Only the first class is eligible by default.

    line 65
  9. Very specific:

    Write the artifacts and machine state defined in references/repo-bundle-contract.md.

    line 96
  10. Very specific:

    Run scripts/validate_style_bundle.py before calling the profile usable.

    line 106
  11. Distinctive:

    Do not rewrite merely to make numeric rates resemble the corpus.

    line 117
  12. Distinctive:

    verify every selected hash against the source PDF.

    line 141
  13. Distinctive:

    when repository rules require it, add the project activity entry and commit/push from the authoritative checkout.

    line 145

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Paper Style Profiler

Build a repository-local, reusable English-language style contract from published full-text papers that fit the current manuscript. Treat the corpus as evidence about language choices, not as text to copy and not as authority for scientific claims. The bundled metrics helper is English-specific; do not apply its lexical metrics to another language without a validated analyzer for that language.

Non-negotiable boundaries

  • Separate language style from argument structure. Subjects, voice, verbs, tense, modality, sentence construction, connectives, punctuation, and paragraph rhythm belong here. Section order, contribution count, method design, experiment design, and figure layout do not.
  • Aggregate three to five complementary papers. Never imitate one author sentence by sentence, reproduce distinctive wording, or make a near paraphrase.
  • Preserve the target manuscript's facts, numbers, polarity, uncertainty, causal strength, scope, baselines, citations, terminology, and locked spans. Style may make a claim clearer; it may not make it stronger.
  • Treat measurements as descriptive evidence, never quotas, acceptance predictors, quality scores, AI scores, or authorship classifications.
  • Do not optimize for detector evasion. If that is requested, decline that objective and offer a conventional academic-prose audit for formulaic, repetitive, vague, or legalistic writing.
  • Use repository papers by default. Browse or add external papers only when the user explicitly requests it or governing instructions require current external verification.

Choose the operating mode

Inspect the repository and any existing style state before choosing a mode:

  • STATUS: report corpus, anchors, hashes, profile age, target coverage, and stale conditions without writing.
  • DISCOVER_AND_EXTRACT: no valid saved profile exists, or the user asks to build one from the repository.
  • REFRESH: candidate papers, selected anchors, venue, target manuscript, or hashes changed.
  • AUDIT: compare manuscript language with the saved profile; do not revise unless requested.
  • WRITE_OR_REVISE: draft or minimally revise specified manuscript text using a valid saved contract.

If the request is ambiguous, prefer AUDIT over mutation. A saved profile does not authorize editing the manuscript.

Recognize existing project style work

Before creating a new bundle, search scoped project locations for an existing corpus contract, language profile, metrics, builder script, or project validator. Treat it as LEGACY_PROFILE_FOUND, not as absent state. It may be adopted only when the selected full texts are identified and hash-pinned, selection rationale is recorded, the artifact covers the actual language dimensions in this skill, source sentences are not reused, and the repository's own validator still passes.

Do not silently rewrite a legacy layout into .paper-style/. In STATUS, report whether it is usable under project governance and which refresh operations are impossible without the original full texts. In REFRESH, require verifiable source files; notes and metrics alone are not a substitute for the corpus. Convert to the standard bundle only when conversion is part of the requested scope.

Resolve authority before analysis

  1. Find and read all applicable repository instructions, especially AGENTS.md or CLAUDE.md.
  2. Identify the authoritative checkout or remote worktree, repository root, branch, HEAD, status, target manuscript, target venue/field, and stated claim ceiling.
  3. Follow repository-specific source-of-truth and commit rules. Do not write to a disposable mirror when the repository declares another checkout authoritative.
  4. Identify the bundle location. Use an existing configured location if present; otherwise propose .paper-style/ and state that path before writing.
  5. Keep raw PDF text in an explicit temporary directory outside the repository. Never place extracted source text in the saved bundle.

If authority or the target manuscript is materially ambiguous, stop with BLOCKED_AUTHORITY_AMBIGUOUS and name the missing decision.

Discovery and extraction workflow

P0 — Resolve the project

Record the authority, target, venue/genre, scientific claim ceiling, and applicable exclusions. Do not infer these from filenames alone.

P1 — Freeze the candidate inventory

Use scoped repository search (rg --files first) and avoid dependencies, caches, generated builds, virtual environments, raw datasets, and unrelated archives. The helper scripts/inventory_papers.py may inventory PDFs and derive hashes and extraction-quality signals without retaining source text. Invoke it as python3 scripts/inventory_papers.py <repo-root> --scope <relative-dir> --output <bundle>/CORPUS-INVENTORY.json (--scope is repeatable; add --include-ignored to walk Git-ignored paths). It discovers .pdf files only, and its page counts and text-quality signals rely on the optional poppler-utils binaries pdfinfo and pdftotext; when they are absent the extraction fields degrade to tool_unavailable rather than failing.

Classify candidates as one of:

  • external published full-text paper;
  • current or prior manuscript by the project authors;
  • slide, report, thesis, review response, or non-paper material;
  • duplicate or incomplete/unreadable source;
  • unresolved.

Only the first class is eligible by default. Freeze the inventory before close language reading so attractive sentences cannot silently change the sampling frame. Follow references/discovery-selection.md.

P2 — Freeze three to five complementary anchors

Score eligible papers using the documented rubric, then select for both fit and complementarity. Cover at least these rhetorical roles when the repository permits:

  • empirical Results and evidence interpretation;
  • method, mechanism, or system description;
  • Discussion, limitations, or claim-boundary language.

A high total score does not override a claim-ceiling conflict. Record every selected anchor, primary role, score, selection basis, hash, and material exclusion. If fewer than three suitable full texts remain, set INSUFFICIENT_REPOSITORY_STYLE_CORPUS; do not pad the corpus with the current manuscript, duplicate versions, or weakly related papers.

P3 — Extract the language layer

Convert selected PDFs to temporary plain text with a deterministic extractor. Hash the original PDFs, note extraction coverage, and exclude references from measurement where feasible. The helper scripts/profile_language.py produces quote-free descriptive metrics; it does not replace close reading. Invoke it as python3 scripts/profile_language.py --paper A1=<tmp>/a1.txt --paper A2=<tmp>/a2.txt ... [--target T=<tmp>/target.txt] --output <bundle>/metrics.json, where each --paper names a selected anchor ID and its temporary plain-text path (for example from pdftotext, which is optional poppler-utils software).

Extract the dimensions in references/language-profile.md, including:

  • grammatical subjects and research agency;
  • active/passive balance and first-person use;
  • action, evidence, interpretation, comparison, and boundary verbs;
  • tense, aspect, modality, and hedging;
  • sentence length, clause chaining, opening patterns, and punctuation;
  • discourse relations and paragraph opening/development/ending rhythm;
  • section-specific language for Introduction, Method, Results, Discussion, and Limitations;
  • numerical comparison, uncertainty, negative-result, and scope-limiting language.

For every qualitative rule, state whether it is shared, role-specific, or paper-specific. Save no long quotations and no reusable source sentences.

P4 — Save the reusable bundle

Write the artifacts and machine state defined in references/repo-bundle-contract.md. The complete bundle contains:

  • STYLE-STATE.json
  • CORPUS-INVENTORY.json
  • SELECTION-REPORT.md
  • LANGUAGE-STYLE-PROFILE.md
  • WRITING-CONTRACT.md
  • metrics.json
  • TARGET-GAP-AUDIT.md when a target audit was performed

Run scripts/validate_style_bundle.py before calling the profile usable. A mechanically valid bundle proves schema and traceability only, not that the selected papers are scientifically optimal or that revised prose is publication-ready.

P5 — Audit the target

Compare the manuscript with the aggregate and section-specific contract. Report concrete gaps, locations, confidence, and likely effect. Distinguish:

  • a genuine language mismatch;
  • a project-required claim boundary;
  • a deliberate terminology choice;
  • an extraction or genre artifact.

Do not rewrite merely to make numeric rates resemble the corpus. Save the audit only when requested or when it is part of a full extraction run.

P6 — Write or revise

Load and verify the saved state first. If anchor or target hashes are stale, refresh or report the stale condition before editing. Apply references/apply-style.md and make evidence-sized changes. For a manuscript-wide task, audit terminology and recurring patterns across the whole document after local edits.

Refresh rules

Refresh rather than silently reuse when any of these changes:

  • an anchor PDF hash or path;
  • the eligible candidate inventory materially;
  • target venue, research genre, or manuscript role;
  • target manuscript identity;
  • selection rationale or claim ceiling;
  • extraction tooling in a way that changes measurements.

A changed target hash alone normally requires a new target audit, not reselection of anchors. Record the reason and retain prior state through version control; do not overwrite provenance invisibly.

Validation and handoff

Before finishing:

  1. validate JSON and the bundle schema;
  2. verify every selected hash against the source PDF;
  3. confirm three to five unique external anchors and their complementary roles;
  4. scan the bundle for raw extracted text, long source quotations, private absolute paths, and unsupported quality claims;
  5. distinguish completed extraction, human judgment still needed, and unverified manuscript quality;
  6. when repository rules require it, add the project activity entry and commit/push from the authoritative checkout.

Report in plain language: what was selected, why those papers, where the contract was saved, what language patterns were found, what was deliberately not inherited, what validation passed, and what remains a human judgment.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.