Claude strip HTML tool examined

This Claude strip HTML tool performs removal rather than parsing. Script and style blocks are deleted in full, contents included. Every remaining tag is removed along with its attributes. Four HTML entities are then decoded. Nothing else happens to the text.

The distinction between removal and parsing matters more than it sounds. A parser builds a document tree and can therefore reason about structure, converting a block element into a paragraph break. A remover matches and deletes. This route is the second kind, which makes it fast, predictable and local, and which also explains every limitation below.

Four passes run in a fixed order, and two of the surprises people report are consequences of that order rather than bugs.

Pass one: script blocks

A script element and everything between its tags is deleted, case-insensitively.

Pass two: style blocks

The same treatment for style elements, so no CSS survives into the text.

Pass three: remaining tags

Anything enclosed in angle brackets is removed, including attributes and ordinary comments.

Pass four: four entities

  becomes a space, & an ampersand, < and > angle brackets. Because this runs last, decoded brackets are not re-stripped.

There is nothing to configure, since the route exposes no options.

Step 1

Paste the markup. Length is unconstrained because the operation is local.

Step 2

Run it. The four passes execute in order, in a single traversal each.

Step 3

Review the first line and the last, which is where head content and unclosed structures surface.

Line breaks are not synthesised. A block-level element does not become a paragraph break, because knowing that a div is block-level requires a model of HTML that a removal pass does not have. Source newlines survive, so well-formatted markup returns readable and minified markup returns as a single line.

Text inside head elements is retained. The tags around a title are removed like any others, and the words between them remain, which places a page title at the top of your result. A comment containing an unescaped angle bracket can leave a fragment, because the match ends at the first closing bracket it finds. Entities beyond the four named are untouched, so numeric references and named entities such as " and © persist as literal text.

Artifacts are complete HTML documents. Copying from one yields a document rather than prose, usually with a style block, structural containers and utility classes attached to every element.

Code blocks are the second path: a request for a component, a landing page or an email template returns markup by definition, and extracting the copy from it is a separate job. A third path is a paste out of a rendered page elsewhere, where the clipboard carries markup the reader never saw.

In a content system, unsanitised markup either renders as unintended layout or is escaped and displayed as literal angle brackets to readers. Neither failure is detected by the person pasting, because the paste looked correct in the editor.

In measurement, markup inflates every count. Tags and attribute values are tokenised as words, so length checks, reading-time estimates and keyword density are all computed against text the reader never sees. Stripping first is what makes those numbers mean anything.

Markdown is not HTML. Asterisks, hashes and bracketed links are a separate syntax and survive this tool untouched, which is what the markdown stripper exists for.

Invisible Unicode is also unrelated. A no-break space written as a literal character rather than as an entity passes through here, and the watermark cleaner is the route that removes it along with thirteen other hidden characters. Both of those, like this one, run locally with no allowance attached.

Editors recovering copy from an Artifact for review as prose, where the markup is irrelevant to the judgement being made. Content and SEO teams extracting body text so that length, keyword placement and readability are measured against what a reader actually receives.

Engineers use it as a crude but effective template audit: strip the structure and read what remains, which exposes placeholder text, duplicated strings and copy that was never written.