write-science-well: author-directed scientific prose revision
An open-format agent skill for researcher-directed scientific prose work. The assessed use revises a supplied passage while retaining its claims, figures, citations and terminology, with proposed edits presented for author acceptance.
L1Task assistance (workflow level, not a quality rating) Source: Repository-documented
Research applicability
A concrete scientific-writing assistance task for papers, theses and proposals, with author oversight rather than new scientific claims generated independently.
Getting started
Editorial, untested guidance: read the core SKILL and relevant reference sections, choose a short non-sensitive passage, state what must not change and request a revision with reversible changes. Check every claim/citation before accepting. Do not authorise in-place edits unless intended.
Maintenance
Pinned primary documentation inspected; ongoing maintenance and host-version compatibility not independently audited.
License
README states MIT for the skill; the source books retain their authors'/publishers' rights. Full dependency/rights audit not performed.
Workflow & human oversight
L1 for one researcher-supplied scientific passage revision. The author defines scope and checks/accepts reversible edits while preserving claims, numbers and citations. Internal editing/scanner rounds do not establish autonomous research; small model-graded smoke tests are author-reported.
Workflow level is not an evidence score; higher is not better evidence.
Problem
Researchers need clearer manuscript passages without silently changing their scientific argument, evidence or terminology.
Approach
The host assistant selects the requested mode and audience, loads relevant style guidance, identifies protected content, revises through single-purpose content/structure/sentence/word/mechanics rounds and supplies paste-ready text plus change/cut lists. The optional prose scanner flags signals for attention rather than deciding correctness.
Outcomes (as reported by the originating team)
The maintainer reports small with/without-skill model-graded tests and author spot checks, calling the numbers smoke tests rather than a measurement. No writing improvement, factual-preservation rate or multilingual effectiveness was independently evaluated here.
Limitations
Documentation review of the pinned README, complete core SKILL and house-style only, not every task-specific reference file. A prohibition on fabrication is an instruction, not a guarantee. English sentence-level rules/scanner are not validated multilingual editing. Maintainer evaluation is small/model-graded; no skill/scanner or manuscript experiment was run.
Reproducibility
Primary README/core SKILL/house-style read and cited with hashes/original access dates; selected core document has API-byte/Git-blob crosschecks from the source-preparation chain. No source software or outcome experiment executed.
Adoption requirements
A compatible host assistant or reader of the open SKILL format, author-supplied text/evidence and explicit editing constraints. Protect confidential manuscripts under the chosen host's policies; author verification remains necessary. The scanner needs a suitable Python environment if used; this review did not install or run it.
Comments
No comments yet.
Add a comment
Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.