Why do Claude and other models wrap output in XML tags?

Anthropic recommends XML tags such as thinking and answer to separate a model's private reasoning from its final response. Other prompt frameworks use tags like context, document or scratchpad for the same structuring purpose. They are convenient for the model but usually need to be removed before showing text to an end user.

Does this handle nested tags?

Yes. The stripper walks the text and matches each opening tag with its corresponding closing tag, so a thinking block that itself contains an example tag is handled correctly. Extraction returns the full inner content including any nested tags.

What is the difference between removing tags and removing their contents?

Removing only the tags keeps the inner text but deletes the angle-bracket markers. Removing tags and contents deletes the entire block — useful for stripping a thinking section the user should never see.

Is my text sent to a server?

No. All parsing runs locally in your browser with JavaScript. Nothing you paste is uploaded, logged or stored, so it is safe to use with confidential prompts and completions.

What happens to text outside the targeted tags?

It is left untouched. Only the tags you name (or all tags, if you leave the field blank) are affected, so surrounding prose, markdown and code remain exactly as written.

What is the LLM XML Tag Stripper?

Strip or extract content from XML-like tags used by Claude and other LLMs such as thinking, answer, context and custom tags. Handles nested tags, runs entirely in your browser, and never uploads your text. It runs free in your browser on Gera Tools, with nothing uploaded.

LLM XML Tag Stripper

Name: LLM XML Tag Stripper
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Get one useful tool a week

Like this tool? Enter your email and we'll send you one genuinely useful Gera tool a week — plus a link to come back to this one. No spam, one-click unsubscribe any time.

LLM responses are often littered with XML-style structure — <thinking>, <answer>, <context>, <scratchpad> and any custom tag your prompt asked for. This tool strips or extracts that structure so you get either clean user-facing text or just the inner content you actually want.

How it works

You paste a response and name the tags. The parser scans the text for matching <tag>...</tag> pairs, correctly matching nested opening and closing tags, and then applies your chosen action:

Strip tags only — keep the inner text, delete the angle-bracket markers.
Strip tags and contents — delete the whole block, including everything inside.
Extract contents — return only the inner text of matching tags, discarding everything else.

Everything runs in your browser. Nothing is uploaded, so confidential prompts stay private.

When you actually need this

Pipeline cleaning. You ask Claude or another model to reason step-by-step in <thinking> tags and then give the final answer in <answer> tags. Your backend needs to store or display only the answer. Without a stripper you have to write a regex every time — and regexes for nested tags get messy fast.

Logging sanitisation. Your internal logs should never store a user’s private reasoning trace. Strip tags and contents to discard <scratchpad> blocks before writing to your database.

Structured extraction. A document-QA pipeline wraps retrieved passages in <context> tags. You want to pull just those passages as clean text. Set the action to Extract and the tag to context.

Worked examples

Scenario 1 — hide reasoning, keep answer. The model returns:

<thinking>I should check whether the sum is correct…</thinking>
<answer>The total is $42.</answer>

Tag: thinking · Action: Strip tags and contents. Output: <answer>The total is $42.</answer>. Then run again with tag answer, action Strip tags only to get: The total is $42.

Scenario 2 — extract multiple context blocks. A RAG response contains several <context> passages scattered through the completion. Tag: context · Action: Extract contents. Output: the inner text of every context block concatenated, discarding all prose between them.

Common mistakes and edge cases

Blank tag field acts on every tag detected — useful for a full-strip pass, but be careful if the model’s answer itself contains angle brackets used as comparison operators (< and >), which the parser may try to match as tags.
Nested same-name tags (<thinking>…<thinking>…</thinking>…</thinking>) are handled correctly; the tool pairs each opener with its matching closer.
Partial blocks — if the model was cut off and a <thinking> was never closed, the tool leaves the trailing text verbatim rather than deleting an unmatched opener.