繁體中文 · Quick start · Locale support · Contributing

An agent skill for editing README files, documentation, release notes, contribution
guides, PRs, issues, UI copy, error messages, and prompts—without treating a
blacklist or detector score as a style guide. Runs on Claude Code and Codex.

Meaning stays intact
Protect facts, numbers, versions, conditions, negation, attribution, causality, commands, links, quotations, and markup.

Edits fit the surface
Apply different standards to a README, tutorial, PR, release note, error message, or policy document.

Locales stay specific
Use regional terminology and false-positive guards instead of a universal replacement list.

Quick start

The skill directory follows the Agent Skills format, so the
same files work in Claude Code and Codex.

Install with Claude Code

Add this repository as a plugin marketplace, then install the plugin:

/plugin marketplace add ting-hong-shieh/polish-open-source-prose
/plugin install polish-open-source-prose

To install the skill without the plugin system, copy the directory instead:

git clone https://github.com/ting-hong-shieh/polish-open-source-prose.git
cp -r polish-open-source-prose/skills/polish-open-source-prose ~/.claude/skills/

Use .claude/skills/ in a project instead of ~/.claude/skills/ to scope the skill to
that repository.

Install with Codex

Invoke $skill-installer and ask:

Install the polish-open-source-prose skill from this repository's
skills/polish-open-source-prose directory:
https://github.com/ting-hong-shieh/polish-open-source-prose/tree/main/skills/polish-open-source-prose

You can also use the skill folder directly during local development:

skills/polish-open-source-prose

Invoke the skill

Claude Code loads the skill automatically when a request matches its description. To
invoke it directly:

/polish-open-source-prose

In Codex:

$polish-open-source-prose

Try one of these requests:

Audit this README and propose only evidence-backed edits.

Localize these release notes for zh-Hant-TW without changing product behavior.

Review this PR description for unsupported claims and lost qualifications.

How it works

Establish the source of truth. Inspect code, tests, configuration, and project
terminology before trusting promotional copy.
Protect semantic constraints. Lock facts, qualifications, identifiers,
quotations, legal text, commands, links, and markup.
Diagnose concrete costs. Revise vagueness, unsupported claims, missing actors,
broken logic, repeated canned structures, and surface or locale mismatches.
Run a semantic diff. Compare subjects, numbers, versions, conditions, negation,
attribution, causality, and ordered steps before delivery.

The skill leaves clear, specific, voice-appropriate prose alone. Passive voice,
parallel lists, fragments, questions, dashes, and polished sentences are not automatic
defects.

What it protects

Area	Examples
Meaning	Subjects, scope, comparisons, conditions, exceptions, uncertainty
Evidence	Numbers, dates, versions, attribution, causal claims
Technical text	Commands, flags, APIs, identifiers, paths, URLs, error strings
Quoted and governed text	Quotations, citations, licenses, policies, security steps
Structure	Headings, anchors, tables, lists, code fences, placeholders, frontmatter
Voice	Deliberate humor, community terms, register, and first-person stance

Locale support

Locale	Status	Coverage
English	Core guidance	Open-source editorial signals and surface rules
Chinese	Core guidance	Chinese editing signals and semantic safeguards
`zh-Hant-TW`	Dedicated locale pack	Taiwan terminology, punctuation, register, and forward cases
Other locales	Foundation only	Shared fidelity workflow; native pack and review still required

The
locale pack contract
defines the evidence, terminology, false-positive, surface, and test requirements for
adding a language-and-region target. It is also designed to become a policy layer for
a future PolyglotGuard checker.

Boundaries

This project does not:

determine whether a human or model wrote a passage;
optimize prose to evade an AI detector;
promise removal of a statistical watermark;
invent metrics, product behavior, user stories, opinions, or personal experience;
claim native support for a locale without a reviewed locale pack;
replace legal, security, or domain review.

For authorship provenance, the skill recommends a signed canonical artifact rather
than treating writing style or a statistical watermark as proof of identity.

Validation

Run all repository checks:

python3 scripts/validate_repo.py

Run the skill checks directly:

python3 skills/polish-open-source-prose/scripts/validate_skill.py

The current corpus contains 43 forward specifications: 17 cases that should remain
unchanged and 26 that should be revised or answered with provenance guidance.
Structural checks catch protected-token drift and corpus errors; native review is
still required to judge real project prose.

Repository layout

.
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── .codex-plugin/plugin.json
├── docs/assets/
├── scripts/validate_repo.py
└── skills/
    └── polish-open-source-prose/
        ├── SKILL.md
        ├── agents/openai.yaml
        ├── references/
        ├── scripts/
        └── tests/

Each host reads its own manifest directory and the shared skills/ tree, so adding an
agent platform does not fork the editorial content.

Repository documentation stays outside the skill directory so it is not loaded as
agent instructions.

Contributing

Read CONTRIBUTING.md before proposing a broad editorial rule or new
locale. Particularly useful contributions include:

false-positive reports;
missing semantic safeguards;
contextual regional terminology;
examples that can be redistributed;
balanced change/keep forward cases;
native review of locale packs.

License and acknowledgments

Original contributions are licensed under Apache-2.0. Material derived from
hardikpandya/stop-slop remains under its MIT license. See
THIRD_PARTY_NOTICES.md and
LICENSE.stop-slop.

The design was informed by public work from
stop-slop,
speak-human-tw,
Humanizer-zh-TW,
humanizer-zh-tw, and
Humanizer-zh-TW-Pro.