CleanLanguage.ai

Disclosure

Disclosure

This page sets out what the project claims, the limit of each claim, and the evidence. Each entry covers 1 thing the project does.

Entry 1

What the skill is made of

Claim: The skill is a zip file of Markdown instructions and references, 1 YAML configuration file, PNG and SVG icons, and the licence and notice files. It contains no program and makes no network calls. The release packager refuses to build the zip if any file in the skill folder is executable or a link, or falls outside its list of allowed paths.

Limitation: The packager’s allowlist step checks each file’s path and mode, not its content. Its other checks include the version, name and release-date lines of SKILL.md, the version and name in the Claude plugin manifest, the name and source of the first plugin the marketplace manifest lists, the icons the YAML file names, the archive’s entry list and layout, and that each archived byte equals its git blob. These are structural checks; none of them judges what the instructions say. The repository that publishes the skill also holds code that is not in the zip. That code includes Python and shell tools, GitHub workflows, the site’s JavaScript, and the AIQT Guardrails hooks used while maintaining the repository. Some of those tools read release data from GitHub.

Evidence: Trust and assurance The release packager (opens in a new tab) Distribution and assurance (opens in a new tab)

Entry 2

What it does to your writing

Claim: The skill tells your AI to remove generic AI-writing patterns and to keep facts, figures, names, quotations, citations, and technical terms unchanged. Its precedence order puts factual, technical, legal, and semantic accuracy first, so a style rule gives way when it would change the meaning.

Limitation: Each provider controls how its model applies the instructions, so results vary from one platform to another. A model can still change meaning with these rules in place, so read the result before you rely on it. For legal, medical, financial, safety, or other high-stakes text, have a qualified person review any edit. The skill does not target AI-detector scores.

Evidence: The precedence order Trust and assurance The skill’s source (opens in a new tab)

Entry 3

Where it works

Claim: The install guide gives steps for Claude, ChatGPT, Gemini, Copilot, and other AI systems that accept custom instructions. Each platform’s page names the date its steps were last checked: 27 August 2026 for Claude, ChatGPT, and Gemini, and 1 September 2026 for Copilot.

Limitation: The steps match each platform as checked on that date, and no check in the repository rechecks them. Skill upload is documented for desktop browsers; the mobile apps are not documented to support it. Whether a platform applies the skill automatically or only when you ask varies by platform and account. The condensed and compact instruction files carry less of the finer detail than the full file.

Evidence: Install Install in Claude Install in ChatGPT Install in Gemini Install in Copilot Install in other AI systems How to use Clean Language

Entry 4

Downloads and checksums

Claim: Each release publishes a SHA-256 checksum for each zip it ships. Once the site is updated after the release, the verify page shows that checksum after a sentence that begins “The published checksum for version” and names the version; until then, it shows no checksum and says that the checksum for the version “appears here once that version is published”. A daily check, tools/check-release-checksum-live.py, reads the verify page as committed in the repository, site/verify/index.html, and compares the checksum it shows with the published release file; it also runs on each change to that page. While a version’s release and its tag do not exist, no checksum for it is published, so the check fails if the page shows one as published; if the check cannot read whether they exist, it could not complete. That check and an offline gate, tools/check-release-links.py, each require the committed verify page to be, byte for byte, the page the site generator writes from its sources: the offline gate for the release record, site-src/release.json, and the daily check for the version and checksum the page shows. The page therefore carries its download steps exactly as the generator writes them, and nowhere else: the checksum and version in that record, or, while the record holds a placeholder, no checksum and 1 note that the checksum is not yet published. A test of the generator pins the whole verify page and the content of any stylesheet of the site’s own, so a change to the page’s text or markup other than its release values and date, or to that stylesheet, fails that test.

Limitation: A match shows that your copy matches the published release file. The instruction files download from this site and have no published checksum. Rebuilding the zip yourself gives the same checksum only with the same zip implementation. In short, before the release of the version the verify page names is published, the daily check can pass only before that version’s release tag exists, on a page that names the current version and says that its checksum is not yet published. The exit-code list at the top of tools/check-release-checksum-live.py states the full condition of a pass and of each other outcome. For example, another version with no release fails, as does a checksum shown as published for a version with no release and no tag, and a malformed Released line, or a tag with no published release, means the check could not complete. The tools that write the release record and the site refuse a checksum that is not the placeholder but has 56 or more of its 64 digits 0, such as a mistyped placeholder; the offline gate fails a release record that holds one, and the daily check refuses a page that shows one, before it reads the release. Between a release and the site update that follows it, the verify page still shows no checksum and says that the checksum for the version “appears here once that version is published”, although the release is published. The daily check reads the page as committed, not as deployed: once the release is published, a run on a commit whose page still says so does not pass, as the check’s exit-code list states, so the next daily run shows a forgotten update. The update is committed before the site deploys it, so a run on the update, or after it is merged, reads the updated page, and the time the deploy takes cannot make the check fail. Neither check reads the meaning of the page’s text. A change made in its sources and regenerated passes the offline gate unless the regenerated page breaks the rule that both checks apply or another check of the offline gate. The daily check stops such a change only if the regenerated page breaks that rule or another check that the daily check makes before it reads the release, such as that the page carries the note, or a checksum and the sentence that names its version. A plain 64 digit hexadecimal value outside the download steps fails both. A zip file name for another version, written in place of the current one, fails the offline gate but not the daily check. The same 64 digit value with its groups joined by dots passes both checks, and the pin stops it. Updating the pin is a decision that review makes. The pin holds the page as written for a stand-in release and a stand-in date, 2 January 2000, and the same test requires 2 pages to be the pinned page with the stand-in release and date replaced: the page that both checks require, written for the release record at the dates in site-src/site.json, and the page written for the same version with a test checksum on 31 December 2099. A generator change that writes differently only for release values and a date that none of these pages uses is not stopped by the pin, that test’s comparison or, once regenerated, either check, until the committed release record and verify page date are values it writes differently for; the test runs on every pull request, so it fails then. Neither check reads the stylesheet and script files the page loads. The verify page loads the 3 stylesheets and 2 scripts of the site template, and tools/check-baseline-pin.py holds each to the file the pinned template version publishes, so an edit to one fails that gate; a new template version changes the page, so it fails the pin until the pin is updated. The site has no stylesheet of its own; adding one fails the pin until the pin is updated.

Evidence: Verify your download The daily checksum check (opens in a new tab) The offline release gate (opens in a new tab) Releases (opens in a new tab)

Entry 5

Data handling

Claim: The skill carries no telemetry and sends nothing to the project; it runs inside your own AI assistant. The site’s own code sets no cookies, and its scripts make no network calls.

Limitation: Your writing goes to the AI provider you use, under that provider’s terms. Cloudflare delivers the site and GitHub serves the skill download; both receive your requests and keep their own records. Cloudflare may set a security cookie, and the security policy of these pages allows Cloudflare’s own scripts, including its Web Analytics beacon, which reports visits to Cloudflare.

Evidence: The privacy policy (opens in a new tab) The site’s security headers (opens in a new tab)

Your browser

If you use the light and dark switch, your choice is kept in your browser’s local storage under the name scheme, so the site opens in that scheme next time. It is not sent to this site, and nothing is stored until you press the switch.

Security

To report a security vulnerability, see the vulnerability disclosure policy (opens in a new tab). For how personal data is handled on this site, see the privacy policy (opens in a new tab).