SkillAtlasSkill 详情

iso-24495-5

Seven Agent Skills that support plain language writing, document audits, code, and organisationa...

审核状态:已审核Quality 80Security 70

复制安装命令

用 Codex 或 Claude 安装复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它先审查 Skill 页面再帮你安装。

复制前请先查看来源、License 和安全提示。

项目 README

来源文件:README.md

抓取于 2026年9月12日

ISO 24495 Plain Language Skills

Seven Agent Skills that support plain language writing, document audits, code, and organisational implementation. They apply principles inspired by the ISO 24495 Plain language series.

The skills are plain SKILL.md files with agent-neutral wording. Any tool that reads the Agent Skills format can use them.

This repository also packages them as a Claude Code plugin with an ISO 24495 output style (output-styles/iso-24495.md). Select the style with /output-style to hold every response to the core rules without relying on skill activation.

Skills

SkillScope
iso-24495-1Core principles. Governs all user-facing output: no filler preambles, short sentences and paragraphs, active voice, scannable structure, concrete instructions.
iso-24495-2Legal writing. Extends the core skill for contracts, licences, and compliance text: standardised modal verbs, no legalese, named actors, structured conditional clauses.
iso-24495-3Science and technical writing. Extends the core skill for documentation, architecture, and code review: progressive disclosure, exact file citations, defined acronyms.
iso-24495-4Organisational implementation (provisional). A task skill for plain language gap analysis in organisations: a process-artefact sweep, a corpus audit, a five-dimension maturity model with deterministic scoring, and an append-only audit trend. Ships TypeScript tooling run with Bun (bun test covered). Based on the unpublished ISO/CD 24495-4 committee draft.
iso-24495-5Document design (provisional). Extends the core skill for structuring complex documents: an opening block, visual hierarchy, navigation aids, layered detail, comparisons, consistent signalling, and a signpost for the reader who wanted a different document. Ships a decision record, a runbook and a design document template. Based on the unpublished ISO/WD 24495-5 working draft.
iso-24495-codePlain language in code. Applies the principles to what a person reads in source: the order units appear in, their names, what comments say, and what an error tells the reader who hits it. Measured to change how Claude structures a file, at no cost to correctness.
iso-24495-text-auditUser-invoked text audit. Checks a selected .md, .markdown, or .txt file or directory. Reports mechanical findings with locations, without deciding validity or compliance.

The core skill activates the relevant writing skills automatically. It triggers iso-24495-2 for legal content, iso-24495-3 for technical content, and iso-24495-5 for complex documents. The text audit never activates automatically.

All skills exempt internal reasoning. The writing skills preserve code blocks, commands, and logs untouched; iso-24495-code is the exception, because governing code is its subject. Technical and legal accuracy always supersede formatting rules.

Installation (Claude Code)

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

/plugin marketplace add https://github.com/GaZmagik/iso-24495.git
/plugin install iso-24495-plain-language@iso-24495

Use the full HTTPS address as shown. The short owner/repo form makes some Claude Code versions clone over SSH, which fails without GitHub SSH keys.

Or from a local clone:

/plugin marketplace add ./path/to/this/repo
/plugin install iso-24495-plain-language@iso-24495

Installation (Codex CLI)

Codex reads the same marketplace manifest, so the plugin installs from the same address:

codex plugin marketplace add https://github.com/GaZmagik/iso-24495.git
codex plugin add iso-24495-plain-language@iso-24495

Or from a local clone, where . is the repository root:

codex plugin marketplace add .
codex plugin add iso-24495-plain-language@iso-24495

Every skill carries agents/openai.yaml, which gives Codex its display name, its short description, and the prompt Codex offers for it. Invoke a skill by name, as in $iso-24495-1, or ask for it in words.

Codex has no output style, so the same rules are a skill there: iso-24495-style holds the output style word for word, and a test keeps the two identical. It lives in codex-skills/ rather than skills/, because Claude Code scans skills/ and would otherwise offer a skill its output style already covers. Codex reads both directories, named in .codex-plugin/plugin.json.

Name the skill in your AGENTS.md to apply it to every response:

Apply `iso-24495-style` to every response.

Put that in your project's AGENTS.md or in ~/.codex/AGENTS.md. An AGENTS.md inside a plugin is ignored, so a plugin cannot apply itself.

Usage

Once installed, the agent loads the skills when their descriptions match the task. To apply one explicitly, ask for it by name, for example: "Apply iso-24495-2 to this licence text."

Invoke iso-24495-text-audit directly and supply one file or directory. The skill reads only that path and leaves every change to the user:

/iso-24495-plain-language:iso-24495-text-audit docs/policy.md

To enforce the core skill on every response, add a line to your agent's instruction file (CLAUDE.md, AGENTS.md, or equivalent):

- ALWAYS activate and adhere to the `iso-24495-1` Plain Language skill across all responses

For agents without a plugin system, copy the skills/ subdirectories into wherever the tool discovers skills.

Disclaimer

This unofficial project is not affiliated with, endorsed by, or approved by the International Organization for Standardization (ISO). The skills contain original guidance inspired by the ISO 24495 series. They do not reproduce the text of any ISO standard.

Publication status: Part 1 published 2023, Part 2 August 2025, Part 3 May 2026. Parts 4 and 5 remain unpublished drafts (ISO/CD 24495-4 and ISO/WD 24495-5). Their skills are provisional guidance from public scope statements, to be revised when ISO publishes.

Conformance disclaimer. The full ISO 24495 texts are licensed and have not been consulted. These skills are built from public principles, published scopes, and common plain-language practice.

The principles derive from the International Plain Language Federation's freely published framework. Every quantitative rule here (sentence length, paragraph density, legalese, heading depth) is this project's own proxy. No rule is a clause of any standard.

Nothing this plugin produces is a statement of ISO conformance. No certification scheme exists for ISO 24495. "Aligned" in the skills means aligned with this project's interpretation, nothing more.

Reference

Read the standard rather than this project's reading of it.

The ISO texts are licensed, so the standards themselves cost money. Everything in this repository is built from the freely published material above.

What the engine reads

A rule can only be as right as the text it reads. So the engine parses Markdown the way CommonMark describes it: each line is matched against the containers already open, then against any container it starts. What remains is the block a rule measures. That is what lets a wrapped list item, a quotation continuing without its marker, and a heading written inside a list all be read correctly.

Measured, because a reader reads them:

  • paragraphs, wherever they sit;
  • list items, which are often the longest sentences in a document;
  • quotations, including GitHub alerts such as > [!WARNING];
  • headings, at any depth and in any container;
  • HTML, because its text is prose a reader reads.

Not measured, because they are not sentences:

  • fenced and indented code, which is a specimen rather than advice to give back to the writer;
  • tables, whose cells belong to a grid, except that table-header reads them;
  • YAML front matter, which is metadata;
  • a GitHub alert label, which is a label;
  • a task marker, which is a control rather than two words.

The parser is checked against the CommonMark reference implementation. 302 documents are recorded in skills/iso-24495-4/tests/fixtures/reference-blocks.ts, and every one that this engine reads differently carries the reason why. The reference is not a dependency: it was installed outside the repository, asked once, and its answers kept.

User-invoked text audit

The iso-24495-text-audit skill audits a selected .md, .markdown, or .txt file or directory. It uses the same rule engine as the Part 4 corpus audit. It reports each finding with its file, line, rule, and explanation.

The rules cover sentence length, sentence averages, paragraph length, legalese, and heading depth. They also cover heading-skip, heading-style, acronym-undefined, doublet, prose-enumeration, link-text, image-alt, wordy-phrase, complex-word, double-negative, filler-opening, and table-header.

The last two serve readers who hear or touch a document rather than look at it. A screen reader can list every link with no sentence around it, and an image without alternative text is silence.

The result reports zero findings when no implemented rule fires. That result does not prove the text suits its audience or purpose.

The shipped acronym list stays universal, so a technical vocabulary needs naming per project. Create .iso-24495-4/acronyms.json with the terms your readers already know:

["SQL", "SDK", "CSS", "IDE"]

An unreadable or malformed file leaves the shipped list alone, because an advisory tool must never be the reason a document cannot be checked.

The skill never runs automatically. It requires Bun and does not alter the selected text.

Directory audits skip selected or nested symbolic links and directory junctions. The result reports each skipped entry instead of reading beyond the selected path or following a cycle.

Testing policy

Run bash scripts/check.sh before you push. That script is the whole gate, and GitHub Actions runs the same file on every pull request. A failure on the server therefore reproduces locally with one command. New checks belong in the script, never in the workflow.

bun test always measures coverage. Every measured source file must cover 100% of lines and functions. Test files are excluded from those totals.

The current suite covers 100% of measured source lines and functions.

Bun reports line and function coverage only in this toolchain. We make no branch-coverage claim.

Logic-free composition roots are separate entry files. Tests never import them, so Bun excludes them from the coverage report. End-to-end tests still exercise those entries.

Every new test receives a mutation check. The implementation is deliberately broken, the test must fail, and the correct behaviour is then restored.

TypeScript style

This project follows the Google TypeScript Style Guide. It uses kebab-case filenames instead of snake_case and double quotes instead of single quotes. Both deviations match the wider ecosystem, and the repository conventions test enforces the mechanically checkable rules.

Why this project holds itself to these rules

This repository is both the tool and a user of the tool. Its shared gate audits every supported document, including this file.

That is deliberate. A plain language project that exempts itself has no claim on anyone else. The Part 4 maturity audit runs against this repository first, and its findings are acted on here first.

Roadmap

All seven skills and the output style ship in v0.6.0. What remains:

  • When ISO publishes Part 4: revise the provisional iso-24495-4 skill against the published text. Its committee-draft text is not public, so the current maturity model is original guidance.
  • When ISO publishes Part 5: revise the provisional iso-24495-5 skill against the published text.

Plain-language checks on script comments were once planned for this release. That plan is cancelled. Comments are fragments, and checking them well would cost more machinery than the advice is worth.

Licence

MIT

开发与工程文档与办公内容与创作

高风险

  • 来源需自行核对维护者身份。
  • 未检测到明显脚本安装指令。
  • 未检测到明显外部权限要求。
  • 存在潜在风险命令,请谨慎安装。
  • 扫描发现:3 条。

Codex — Git Clone 安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 克隆仓库:git clone https://github.com/GaZmagik/iso-24495.git
  3. 将 "skills/iso-24495-5" 文件夹复制到 Codex 的 skills 目录中。
  4. 重启 Codex 让新的 skill 生效。

Codex — 手动复制安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 从源仓库下载 SKILL.md 及相关文件。
  3. 在 Codex 的 skills 目录中创建新文件夹。
  4. 将所有 skill 文件复制到新文件夹中。
  5. 重启 Codex 让新的 skill 生效。

Claude Code — Git Clone 安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 克隆仓库:git clone https://github.com/GaZmagik/iso-24495.git
  3. 将 "skills/iso-24495-5" 文件夹复制到 Claude Code 的 skills 目录中。
  4. 重启 Claude Code 让新的 skill 生效。

Claude Code — 手动复制安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 从源仓库下载 SKILL.md 及相关文件。
  3. 在 Claude Code 的 skills 目录中创建新文件夹。
  4. 将所有 skill 文件复制到新文件夹中。
  5. 重启 Claude Code 让新的 skill 生效。

Cursor — Git Clone 安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 克隆仓库:git clone https://github.com/GaZmagik/iso-24495.git
  3. 将 "skills/iso-24495-5" 文件夹复制到 Cursor 的 skills 目录中。
  4. 重启 Cursor 让新的 skill 生效。

Cursor — 手动复制安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 从源仓库下载 SKILL.md 及相关文件。
  3. 在 Cursor 的 skills 目录中创建新文件夹。
  4. 将所有 skill 文件复制到新文件夹中。
  5. 重启 Cursor 让新的 skill 生效。

GitHub Copilot — Git Clone 安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 克隆仓库:git clone https://github.com/GaZmagik/iso-24495.git
  3. 将 "skills/iso-24495-5" 文件夹复制到 GitHub Copilot 的 skills 目录中。
  4. 重启 GitHub Copilot 让新的 skill 生效。

GitHub Copilot — 手动复制安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 从源仓库下载 SKILL.md 及相关文件。
  3. 在 GitHub Copilot 的 skills 目录中创建新文件夹。
  4. 将所有 skill 文件复制到新文件夹中。
  5. 重启 GitHub Copilot 让新的 skill 生效。

Windsurf — Git Clone 安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 克隆仓库:git clone https://github.com/GaZmagik/iso-24495.git
  3. 将 "skills/iso-24495-5" 文件夹复制到 Windsurf 的 skills 目录中。
  4. 重启 Windsurf 让新的 skill 生效。

Windsurf — 手动复制安装

  1. 安装前请先查看来源仓库和风险报告。
  2. 从源仓库下载 SKILL.md 及相关文件。
  3. 在 Windsurf 的 skills 目录中创建新文件夹。
  4. 将所有 skill 文件复制到新文件夹中。
  5. 重启 Windsurf 让新的 skill 生效。
查看 SKILL.md 原文
name: iso-24495-5
description: Provisional sector-specific Plain Language standard for document design (based on ISO/WD 24495-5, under development). Applied when structuring complex documents so readers can find and navigate content through layout, visual hierarchy, and navigation aids.
metadata:
  version: "0.6.2"
  iso-standard: "ISO/WD 24495-5"
  iso-status: "working-draft"

ISO/WD 24495-5 - Plain Language (Document Design) [PROVISIONAL DRAFT]

Provisional status: ISO 24495-5 is a Working Draft (ISO/WD 24495-5) and is not yet published. This skill is original guidance based on the draft's public scope and established information design practice. It does not reproduce ISO text. Expect revision when the standard is published.

Sources: several rules here paraphrase the Document design pattern library, version 0.6, June 2025. That library is by Waller, van der Waarde, Schriver, Slabbert, Cheek and Linsky, for the International Plain Language Federation. The wording in this skill is ours, and no substantial wording is copied from it. Read the Document design pattern library at the International Plain Language Federation for the original.

Extends ISO 24495-1:2023 for the structural design of complex documents: reports, specifications, guides, contracts presented as documents, and long-form technical or health information. Design works together with linguistic cues to help readers find and navigate a document's structure and content.

Design for readers who are not looking at the page. The intended readers include everyone who uses the document. Some see it, some hear it through a screen reader, and some read it by touch.

A listener has no visual hierarchy. Their structure is the heading tree, the link text and the reading order. Every rule below is written to hold when the document is heard.

Scope & Execution Boundaries

  1. Thinking Block Exemption:

    • Internal layout planning and structural reasoning within thinking blocks (<thought>, <thinking>) are 100% exempt from these constraints.
    • Plan freely within thinking blocks. Apply document design rules strictly to final user-facing documents.
  2. Design as Engineering, Not Decoration:

    • Base every design decision on a documented reader need (finding, navigating, comparing, acting). Never add visual elements for aesthetic effect alone.
  3. Content Primacy:

    • Document design must never cut or distort content to fit a layout. Accuracy and completeness supersede visual tidiness.

Required Templates

Read the matching template before writing any of these document types:

  • Architecture decision record (ADR): Read the template file at assets/adr-template.md.
  • Runbook: Read the template file at assets/runbook-template.md.
  • Design document: Read the template file at assets/design-doc-template.md.

Restructure an Existing Document

When asked to restructure an existing document:

  1. Identify the reader tasks, current hierarchy, and navigation needs.
  2. Preserve every prose passage and content item.
  3. Change headings, list types, table structure and visual formatting. You may also move a whole sentence or block, unchanged. Move it only where its dependencies, its order against neighbouring steps, and the claim it qualifies all survive.
  4. Build the opening block, the overview and the signposts from sentences the document already holds, or from wording the author gives you in the request. Promote a sentence only when it states the field directly, and stays both true and complete enough for that field once away from the paragraph it came from. Where the author's wording contradicts the document, promote neither and report the mismatch.
  5. Where nothing serves, leave a marked slot such as [Author needed: purpose] and report it as a gap. This covers the overview's content as much as its label. Never supply the missing wording yourself.
  6. Check the result against the hierarchy, navigation, structure, and signalling rules below.

Do not rewrite prose, change tone, or remove content. Those changes belong to Parts 1 to 3, and so does rewording a sentence to make it fit a slot. A wrong purpose sends a reader confidently in the wrong direction, which is worse than no purpose at all.


Quantitative Rules & Hard Constraints (User-Facing Documents)

  1. Visual Hierarchy Limits:

    • Use at most 3 heading levels below the document title. Flatten deeper nesting into lists or tables.
    • Make headings state the section's message or task, not just its topic ("Install the dependencies" rather than "Dependencies").
    • Two exceptions, and no others. A document type with a published structure keeps that structure's section names, as a decision record keeps Context and Decision.
    • A reference section a reader jumps to by subject keeps the subject as its name, as a specification keeps Data Model. A section read in sequence gets a message or a task.
    • Reject a heading that jokes, puns or plays with words. Reject one built on a term the document has not yet explained, because a reader skimming meets the heading first.
  2. Navigation Aids:

    • Add a table of contents or link list to any document with 6 or more sections.
    • Keep heading wording identical between the table of contents and the section it points to.
    • Number headings when a reader must cite one by its identifier, and never below the depth limit above.
  3. Chunking & White Space:

    • Present one idea per visual chunk (paragraph, list, table, or callout). Separate chunks with blank lines.
    • Never run two unrelated topics together in one paragraph or one table.
  4. Choosing the Right Structure:

    • Comparisons: Use a table when readers must compare 2 or more items across shared attributes. Name the narrowest presentation the table must survive, then read it back at that width. Where nobody has named one, use repeated labelled records instead of a table, rather than shipping both.
    • Sequences: Use a numbered list for steps that must happen in order. Keep it an ordered list rather than numbers typed into a paragraph, so the sequence survives when the document is heard.
    • Options and collections: Use a bulleted list for unordered sets of 3 or more items. Keep each bullet to one paragraph carrying one idea, and nest no deeper than 2 levels. Promote longer material to a subsection.
    • Branching routes: When a procedure forks, use a decision table or a labelled set of conditions rather than one numbered list. A decision table with labelled routes is already the written form. Add prose only where the routes are drawn as a picture.
    • Warnings and conditions: Reserve a callout for a warning or condition that changes what the reader does. Merge adjacent callouts serving one purpose, and give each a word naming what it is.
  5. Consistent Visual Signalling:

    • Give each visual device (bold, italics, blockquotes, code formatting, icons) one meaning per document and apply it consistently.
    • Never use the same device for two different meanings, or two devices for the same meaning. A text alternative is exempt only where the original is a picture or a diagram, or is not exposed to a screen reader. A table is excluded only where its headers identify every value and its reading order keeps the comparison intact. Where they do not, a concise equivalent is allowed.
    • Never let a visual device carry meaning on its own. Bold, colour, an icon and a position on the page are all silent to a listener. State the meaning in words as well. "Required fields are marked in red" fails; "Required fields are marked with the word required" works.
  6. Reaching Readers Who Cannot See the Page:

    • Link text names its destination. A screen reader can list every link in a document, read aloud without the sentence around it. "Click here" and a bare web address tell that reader nothing.
    • Every image that carries meaning has alternative text describing what it shows, not what it is. An image that carries no meaning is decorative and may say so.
    • Tables carry a header row, because a listener hears each cell announced against its column name.
    • The reading order is the document order. A sidebar or a floating callout only makes sense out of sequence, so give each one its own heading in the flow.
  7. The Opening Block:

    • Open every document with its title, a one-line statement of its purpose, and its version or date.
    • Name the intended reader in that block. Part 1 decides who that reader is; this rule decides where the answer appears.
    • Give each field its minimum. Purpose states the reader's task and the document's scope. The reader line names the primary audience. The referral names the alternative and when to use it.
    • Where the document cites this skill or ISO 24495-5, say in the document that the standard is an unpublished draft.
  8. Layering the Detail:

    • Label the overview explicitly in any document with 6 or more sections, or one whose conclusion readers need before the detail. A reader who stops at the overview then knows what they hold.
    • That label is the section's heading, and it names the section rather than its message. This rule overrides the heading rule for that one heading, and for no other. It keeps the document's conclusion, the action required, and any essential qualification.
    • Give that label a heading or a word, never a visual treatment alone.
    • Move detail that only some readers need into footnotes, an appendix, or a collapsible block, and keep it reachable from the main path.
    • Use at most 3 levels: overview, main body, and optional detail. Part 3 governs how a technical explanation is worded across them.
  9. Readers Who Have the Wrong Document:

    • Tell a reader who needs something else where to go. Link the related documents, the other language versions, or a person to ask.
    • Put that signpost where a reader will look on realising the document is wrong. Near the top works, or at the end of the opening section.
    • Leave it out when no alternative exists, rather than shipping an empty heading.

Contrastive Examples

Example 1: Structuring Comparative Information

  • ❌ Not aligned (Buried in Prose):
    The Basic plan costs £5 per month and includes 10 GB of storage but no
    priority support, whereas the Pro plan is £15 per month with 100 GB and
    priority support, and the Team plan, at £40 per month, offers 1 TB,
    priority support, and audit logs.
    
  • ✅ ISO 24495-5 (Draft) Aligned:

    Choose a plan based on storage and support needs:

    PlanPrice / monthStoragePriority supportAudit logs
    Basic£510 GBNoNo
    Pro£15100 GBYesNo
    Team£401 TBYesYes

Pre-Output Self-Audit Checklist

Before outputting a complex document, audit against these checks:

  • Hierarchy depth: Are there 3 or fewer heading levels below the title?
  • Heading quality: Does each heading state its section's message, or a name its genre expects, free of wordplay and of terms not yet explained?
  • Navigation: Does a document with 6 or more sections carry a table of contents worded identically to its headings?
  • Numbering: Are headings numbered only where a reader must cite one by its identifier?
  • Chunking: Does each chunk carry one idea, separated from the next by a blank line?
  • Structure fit: Are sequences in ordered lists, sets in bullets, and forks in a decision table or labelled conditions?
  • Comparisons: Is the table tested at a named target width, or are labelled records used because no width is named?
  • Restraint: Is every bullet one paragraph on one idea, nested no deeper than 2 levels, with longer material promoted to a subsection?
  • Callouts: Does each change what the reader does, with adjacent ones merged and each named in a word?
  • Signal consistency: Does each device carry one meaning, no two devices carry the same meaning, and no meaning ride on a device alone?
  • Alternatives: Is a text alternative present only beside a picture or diagram, or beside a table whose headers miss values or whose order breaks the comparison?
  • Links and images: Does link text name its destination, does a meaningful image say what it shows, and is a decorative one marked as decorative?
  • Tables heard: Does every table carry a header row, with headers that identify each value beneath them?
  • Reading order: Does document order match reading order, with each sidebar and displaced callout given its own heading?
  • Opening block: Does the document open with its title, a one-line purpose naming the reader's task and scope, the primary audience, and a version or date?
  • Overview contents: Where needed, does it keep the conclusion, the required action and every essential qualification?
  • Overview label and detail: Is it labelled in words, and has the detail moved to footnotes, an appendix or a collapsible block?
  • Levels: Are there 3 or fewer levels of detail, and is the optional detail still reachable?
  • Signposting: Is the referral near the top or ending the opening section, naming its destination and when to use it, and absent where nothing else exists?
  • Preserved: On a restructure, was every prose passage and content item kept, with no prose rewritten and no tone changed?
  • Safe moves: Did every moved sentence or block keep its dependencies, its order against neighbouring steps, and the claim it qualifies?
  • Never invented: Did every promoted sentence come unchanged from the document or the author, state its field directly, stay true and complete, with mismatches reported and gaps marked?
  • Content primacy: Did accuracy and completeness survive every structural choice, with nothing cut or distorted to fit a layout?
  • Evidence over aesthetics: Does every design element serve a reader need?
  • Provisional label: Where the document cites Part 5, does it say the standard is an unpublished draft?

发现问题?提交给管理员复核

评分:

评论 (0)

暂无评论,成为第一个评论者吧!