AI Edit Approval, Line by Line
A writing plugin that turns AI polishing into individually approvable text patches, preventing silent full-draft rewrites and hidden-character contamination.
When writers hand a nearly finished article to AI for polish, they usually want only a few better turns of phrase, not a passage silently replaced wholesale. For contracts, submissions, and public statements especially, they need to know where every surviving change came from.
The editor plugin reads only the paragraph the user selects and requires the model to return suggestions as patches. Original and proposed text appear side by side, with deletions, replacements, and additions highlighted in separate blocks. Writers can accept or reject one sentence at a time, or edit manually before requesting the next suggestion. The AI can never overwrite the entire draft directly.
After each acceptance, the plugin scans for invisible Unicode characters, anomalous formatting, and control characters introduced through copying. The change log retains the original sentence, acceptance time, and final text, while export can produce a clean version. When an editor or legal reviewer asks, the writer can account for each change they approved.
An early version would provide a patch panel for browser writing contexts and common document editors, supporting polishing, shortening, and tone changes. It does not judge facts for the writer or replace collaborative approval; it simply turns AI editing into a sequence of visible, reversible text choices.
Why now
On August 16, an article argued that Claude’s text watermark could influence wording choices, making authorship of polished language a concrete concern. S2 As of August 17 at 00:33 UTC, the post ranked 10th, with 121 points and 99 comments. S1
Target user
The core user is preparing to submit a contract, manuscript, or public statement. At this stage, the text is nearly final, and its facts and positions usually cannot be rearranged. They want AI to refine only a few phrases, but must verify that every word still reflects the intended meaning. When editors, legal teams, or clients ask where changes came from, they need a record of line-by-line approval rather than an unexplained rewritten draft.
Minimal entry point
The browser extension first reads the current selection and stores a hash of the original text. The model must return structured patches containing an anchor snippet, replacement text, and a brief rationale. The frontend uses text diffs to split patches into independently approvable sentence blocks. Before writing, it verifies the selection and hash again to prevent misalignment after the document changes. A separate control-character scan reports only actual code points and formatting anomalies. The first release covers standard web text fields, then integrates with Word’s revision APIs. Google Docs' API for writing suggestions remains in developer preview and can be supported later. S3S4
Punching above its weight
The first users can come directly from the writers and developers in this Hacker News discussion. S1 Build a no-login selection-diff demo that lets users see a full-paragraph rewrite broken into individual choices. Plugin-store listings should foreground before-and-after views of contract clauses and public statements. Then publish technical posts using anonymized edit samples to explain which characters are cleaned up and which watermarks cannot be detected.
Competitors & gaps
- Microsoft Word Track Changes and Document InspectorGoogle
- Word already marks insertions and deletions and lets reviewers accept or reject them one at a time. Its Document Inspector can also find hidden text and other information. S4 These capabilities work well for human collaboration and can handle externally generated revisions. The gap is that AI requests happen in a separate workflow: users still have to copy a passage, prompt a model, and paste the result back. If the model returns a full rewrite, Word can show the resulting diff but cannot constrain the scope of changes. Nor does it preserve the relationship among the prompt, original selection, and model patch. This product can connect pre-generation constraints with post-generation approval, tying every write to a text block the user explicitly accepted.
- Google Docs Suggesting modeGoogle
- Google Docs Suggesting mode preserves edits as pending changes. Its official API can read suggestions inline, while creating suggestions and accepting or rejecting suggestion threads remain in developer preview. S3 That gives a plugin a document-native interface to build on. Existing Suggesting mode primarily addresses how collaborators submit edits; it does not require a model to return only the smallest patches within a selection. Users can still paste a fully rewritten passage into a document and then inspect each difference. Suggestion threads are not a complete record of AI revisions, either. The product opportunity is to retain the request, original sentence, patch, and final written result. Even if the platform API fails, unapproved text must not enter the body of the document.
How it makes money
Monthly per-seat subscription. The free tier includes basic sentence-by-sentence approval and local undo. Paid tiers add cross-document history, audit exports, and team policies. Users can supply their own model keys so inference costs do not consume the subscription margin.
The case against
An invisible-character scan can easily be mistaken for watermark detection. Claude’s described approach does not add hidden characters; it leaves statistical signals through word choice. S2 Removing control characters can therefore address copy-and-paste contamination, but cannot verify or remove this kind of watermark. Footnotes, links, and comments in rich text also make patch anchoring fragile. Writing to the wrong location after a selection changes could directly damage a contract or statement. Models may also fail to report changes; a structured response does not make its content trustworthy. Retaining original sentences and acceptance times accumulates sensitive text, creating encryption, retention, and deletion costs. If the product cannot clearly distinguish character hygiene from watermark assessment, it will quickly lose users' trust.