Text Hygiene

Inspect hidden Unicode formatting, clean accidental characters, and review prose for vague or formulaic patterns. Every check is deterministic and runs in this browser tab.

Private by default: your text is never uploaded, logged, or stored. This static page contains no model API and sends no content to a server.

What this tool is not: it is not an AI detector and it does not remove or test proprietary model-level watermarks. A style flag is not proof of authorship. Preserve any disclosure or provenance information required by your school, employer, publisher, or law.

Inspect and improve text

Ready. Paste text and choose Analyze text.

Invisible characters

Characters worth inspecting, normalizing, or preserving.

Writing observations

Rule-based prompts for human review, not an authorship score.

Invisible-character report

No tracked invisible characters found.

Writing-quality review

Based on observable style patterns such as filler, vague attribution, inflated wording, chatbot scaffolding, and repeated em dashes. Context matters; review rather than blindly accept.

No tracked style patterns found.

Reviewable output

Output text

Cleanup preserves emoji joiners, variation selectors, and script-shaping joiners by default. Writing edits only shorten a small set of exact filler phrases; they do not invent facts.

Capabilities and limits

Unicode inspection: reports zero-width characters, special spaces, bidirectional controls, soft hyphens, joiners, mathematical invisibles, and variation selectors.
Conservative cleanup: removes common accidental controls, normalizes special spaces, and preserves characters that can change emoji, mathematical notation, or complex-script spelling.
Style review: flags selected patterns described in Wikipedia's observational field guide. Those patterns also occur in human writing and are not evidence by themselves.
No watermark bypass: model-level statistical watermarks are encoded through token-selection patterns, not necessarily hidden Unicode. This tool cannot identify or remove them.
No C2PA removal: file provenance metadata is outside this text-only tool's scope.

Method and sources

The writing review adapts a small, transparent set of observations from Wikipedia's “Signs of AI writing” field guide. Wikipedia itself warns that these are descriptive clues, not rules or reliable proof. Unicode names and behavior follow the Unicode Standard.

Anthropic's public documentation says supported Claude models use an imperceptible, model-level text watermark and C2PA metadata for supported files, but it has not published its detector or enough technical detail for independent implementation. This page therefore makes no Claude-removal or detection claim.