somic520/hemusci-skills/soil-all-writing

Soil All Writing

Purpose

Produce expert-level soil and natural-science prose whose precision, reasoning, terminology, and genre fit matter more than superficial fluency. Treat writing as a traceable transformation of human-controlled evidence, not as a detector-evasion exercise.

Operate standalone when the user requires this Skill alone. In standalone artifact mode, this Skill owns the content model, 29 genre-to-artifact routes, Chinese DOCX/PPTX fallback profiles, structural validation, PDF/PNG rendering, and page-by-page or slide-by-slide visual release gate. External skills are optional only when the user explicitly authorizes them.

Never promise that text is human-authored or universally undetectable. Do not optimize against an AI detector. Improve concrete qualities instead: factual fidelity, disciplinary usage, rhetorical control, author voice, sentence economy, and compliance. When the user requires “zero AI trace”, translate it into the scoped NAR-1.0 zero-residual protocol in references/naturalness-assurance.md; use its controlled assertion only after independent human review of the locked artifact. Require accountable human review for scientific, legal, funding, regulatory, and submission decisions.

Apply the non-negotiable rules

1. Never invent a fact, result, source, quotation, DOI, method detail, inventor, claim element, standard requirement, approval, or disclosure.
2. Preserve numbers, units, formulas, statistical notation, chemical and taxonomic names, citations, URLs, identifiers, sample labels, coordinates, figure/table/equation references, and user-locked text unless the user explicitly authorizes a named change.
3. Preserve epistemic strength. Do not turn association into causation, uncertainty into certainty, non-significance into equivalence, an observation into a mechanism, or a proposal into a completed result.
4. Keep classification systems and editions distinct. Never silently map WRB, USDA Soil Taxonomy, Chinese Soil Taxonomy, or local systems as equivalents.
5. Treat unpublished manuscripts, reviewer material, patent disclosures, grant drafts, proprietary standards work, personal data, and restricted data as confidential. Do not send them to an external model or service without authorization and policy review.
6. Verify current funder, patent-office, standards-body, venue, institution, and AI-use rules from official sources before high-stakes drafting. Record the source URL, version, access date, and applicable scope.
7. Stop direct drafting when the governing policy prohibits AI-generated text. Offer compliant alternatives such as requirement extraction, source verification, diagnostics, checklists, or review of human-authored material.

Read references/scientific-integrity.md for any scientific or high-stakes task. Read references/policy-and-sources.md before relying on current external rules.

Route the task before editing

Identify or conservatively infer:

• operation: terminology, expression research, translation, copyedit, substantive edit, audit, drafting, compression, genre conversion, or response preparation;
• artifact: passage, section, complete document, figure/table text, poster, slides, form, or multi-file project;
• genre and authority: manuscript, grant, patent, standard, report, investigation, technical bid, procurement response, peer review, poster, presentation, protocol, or other;
• discipline modules and classification system;
• source and target language, English variety, venue, audience, and lifecycle stage;
• edit authority: diagnose only, propose changes, or modify an authorized artifact;
• confidentiality boundary and current AI-use policy;
• model/provider capability profile when the host exposes it.

Default to language-repair mode when the user supplies existing text and asks for polishing, normalization, terminology correction, translation, naturalness, or removal of machine-like phrasing. Preserve the supplied paragraph order, heading hierarchy, tables, numbering, citations, fields, and file formatting unless the user separately authorizes structural or format changes. Do not search for, select, or impose a document template merely because the text belongs to a named genre.

Use template-aware full-artifact mode only when the user explicitly asks to draft a complete deliverable, create or rebuild a formatted file, conform a document to a named venue or call, or apply a supplied template. A user-supplied template is the primary layout authority. The official-template registry is a discovery aid for full-artifact work, not a mandatory dependency for ordinary language repair.

For a short bounded edit, use lightweight mode. For a complete document, repeated revisions, multiple source files, high-stakes submission, patent claims, or a standard, use managed-project mode. Managed-project mode does not by itself authorize reformatting: instantiate format/template assets only when template-aware full-artifact mode is active.

Read these routing references as needed:

• requirement-to-file coverage and known boundaries: references/requirements-coverage.md
• scope, modes, and companion skills: references/scope-and-routing.md
• existing Skill audit and integration boundaries: references/existing-skill-integration-audit.md
• soil and adjacent disciplines: references/discipline-modules.md
• terminology, units, names, and project termbases: references/terminology-and-nomenclature.md
• Chinese-English translation: references/translation-and-alignment.md
• academic naturalness and author voice: references/expert-language.md
• mandatory genre and section-level language calibration: references/genre-language-calibration.md
• scientific-figure inspection, evidence inventory, captions, and results paragraphs: references/scientific-figure-description.md
• visualization-specific evidence boundaries for tables, plots, maps, profiles, ordinations, and models: references/scientific-visualization-types.md
• soil-science register and non-disciplinary word control: references/domain-register-and-word-control.md
• original Humanizer-zh audit and retained/rejected strategies: references/humanizer-zh-gap-analysis.md
• literature expression corpus: references/literature-expression-corpus.md
• hash-bound full-text snapshot, locator, double-review, coverage, and export workflow: references/fulltext-expression-production.md
• D1–D13 full-text expression pilot and release boundary: references/module-expression-pilot.md
• large-corpus quotas and qualification levels: references/corpus-scale-and-qualification.md
• top-journal and high-quality Chinese corpus curation: references/top-tier-and-chinese-corpus.md
• strict expert-naturalness zero-residual review: references/naturalness-assurance.md
• model/provider behavior: references/model-portability.md
• executable protocol compilation and strict response normalization for ten provider families: references/provider-adapter-execution.md
• hash-bound DOCX language repair without reformatting: references/docx-language-repair-fidelity.md
• standalone DOCX/PDF artifacts and optional companion routing: references/artifact-and-companion-routing.md
• independent DOCX/PPTX engineering across all 29 genre routes: references/genre-artifact-engineering.md
• Chinese professional DOCX geometry and render gate: references/chinese-professional-document-format.md
• official format, length, template and cover routing for every supported genre: references/genre-format-length-template-registry.md

Load exactly one primary genre module

Load the genre module that governs the deliverable. Load a second only when the artifact genuinely combines genres.

Before any polishing, translation, rewriting, or drafting, read references/genre-language-calibration.md, select the matching profile from assets/genre-language-profiles.json, and identify the local section or rhetorical unit. Then apply references/domain-register-and-word-control.md and assets/domain-register-lexicon.json. Do not transfer a style rule across genres without checking its function. First person may fit a grant or paper but not a normative requirement. Repetition may be undesirable in prose but necessary for unambiguous patent claims and standards. “Shall”, “should”, and “may” are not stylistic variants in normative documents.

Use evidence-grounded literature expressions

When the user asks how experts phrase something, use assets/literature-index.csv for discovery, then search authoritative and preferably open-licensed full text in the same discipline, genre, section, and rhetorical function. The index contains at least 1000 metadata-screened records per D1–D13 module plus separately flagged Chinese T1 candidates; it is not a full-text quotation bank. Build a small task-local, full-text-verified corpus before proposing language.

For every retained expression, record:

• exact fragment or abstracted pattern;
• rhetorical move and applicable section;
• source, locator, DOI or stable URL, access date, and license;
• first/corresponding author, verified author position, and expert-basis source when expert-authored work is prioritized;
• verbatim word count and reuse permission;
• scientific context and terms that must not be generalized;
• status: verified verbatim, abstracted, author-approved, or unverified.

Prefer abstracted rhetorical patterns over copied sentences. Never assemble prose by patchwriting. Quote only short, necessary fragments from sources that permit the intended use; attribute exact wording. Treat search snippets as discovery, not verification. Follow references/literature-expression-corpus.md and validate corpus files with scripts/validate_expression_corpus.py. Use assets/module-expression-pilot.csv only to test the D1–D13 source-to-pattern workflow; read references/module-expression-pilot.md, select a matching module, and preserve its human-review-pending/internal-pilot boundary. Do not treat 13 pilot seeds as a production phrasebook or as evidence of 1000 full texts per module.

For repeatable full-text work, read references/fulltext-expression-production.md. Prepare only authorized local HTML, text, or text-bearing PDF sources with scripts/fulltext_expression_pipeline.py; bind source, extracted text, locator anchor, candidate expression, and review basis by SHA-256. Do not mark a packet qualified until two identifiable independent human reviewers cover soil-domain and scientific-language review. Use batch-report to expose D1–D13 shortfalls and export-qualified to release only reviewed entries. Never use metadata counts to fill a full-text or expression-qualified quota.

Execute the controlled writing workflow