Llama strip HTML pass

The operation is a fixed chain of six replacements applied in order. Script elements are removed with their contents, style elements are removed with their contents, anything else matching a tag shape is removed, and then four entity substitutions run.

Order matters and is worth stating, because it produces two behaviours a user would not predict. The script and style passes run first, which is why their contents are removed rather than orphaned. The entity passes run last, which is why escaped markup such as an encoded div tag is decoded into literal angle brackets after tag removal has already finished, and is therefore not stripped a second time.

There is no configuration surface. No whitelist, no attribute handling, no comment toggle and no line break setting. Execution is local, with no transmission, no storage, no length ceiling and no usage allowance.

Entity handling is where tools in this category diverge most and describe themselves least precisely. The coverage here is small, fixed and fully enumerable, so it is enumerated.

 

Replaced with an ordinary space.

&, <, >

Replaced with ampersand, left angle bracket and right angle bracket.

Named entities outside that set

", ©, —, ’ and the rest are returned unchanged as literal text.

Numeric entities

Decimal and hexadecimal forms such as ' are not decoded at all.

Double encoding

Sequential replacement means < resolves to a bare angle bracket rather than to <. A single pass over correctly encoded input is the supported case.

Three actions, no configuration, and the input is retained throughout for comparison against the result.

Step 1

Paste the markup. No character ceiling applies, because nothing is transmitted and no request is sized.

Step 2

Run the pass. The output field populates immediately and the input field is unchanged.

Step 3

Inspect rather than copy blind, specifically for residual entity text and for head content at the head of the output.

A tag is matched as an angle bracket, a run of characters that are not angle brackets, and a closing angle bracket. That definition is correct for the overwhelming majority of real markup and provably wrong for two cases, both verified against this engine rather than assumed.

Comparison operator in an attribute

A title or data attribute containing a greater-than sign terminates the match at that character, so the remainder of the attribute and the closing bracket are emitted as text.

Angle bracket in a comment

A comment containing the same character leaks its remainder into the output for the same reason.

Malformed markup

No document is constructed, so unclosed elements and invalid nesting are not repaired and are not detected.

Scope statement

This pass is correct for controlled, well-formed content markup. For arbitrary untrusted HTML a parser is the right instrument and this is not one.

Whitespace is never modified. Every newline, indent and blank line in the source survives into the output verbatim, and there is no option to collapse them, because no whitespace operation exists in this chain at all. Competing tools expose this as a checkbox. Here it is a fixed property.

Head content is likewise not special-cased. Some strippers discard the head block entirely. This one treats its elements like any other tag, so the text of a document title survives the pass and appears at the start of the output, typically adjacent to the first body text with no separator.

Both behaviours are consequences of the same design decision: remove tags, change nothing else. Predictability is the benefit, and the cost is that the result of pasting a complete document needs a moment of manual tidying.

Method guides for this category converge on one warning: a single tag-shaped replacement removes script and style tags while leaving everything between them, so executable code and stylesheet declarations end up embedded in what the user believes is prose.

That failure is silent in a long document. The output looks plausible at the top and the bottom, and the defect sits in the middle where nobody reads on a quick check.

The two dedicated block replacements here run before the general pass specifically to prevent it. It is the one respect in which this implementation is strictly better than the naive version of itself, and it is worth stating explicitly rather than leaving as an implementation detail nobody sees.

Queries pairing a model name with tag stripping generally originate in one task: converting retrieved web content into clean input. Markup consumes length while contributing nothing about the prose, which makes removal an obvious first step.

The limitation to hold in mind is that removal is not extraction. A stripped full page still contains navigation labels, footer link text and the document title, all now indistinguishable from body copy. Selecting the region before stripping produces a materially better result than stripping and then selecting.

Editors recovering body copy from a content field that stores raw markup. Teams reducing an email template to its text before rewriting it. Anyone converting a saved article into plain input.

The precondition is that the markup is disposable. Where structure carries meaning this is the wrong instrument, because nothing in the chain translates a tag into a heading, a list or a link. It removes, and that is the entirety of the contract.