Command Palette

Search for a command to run...

Proofreading Without Uploading: What an On-Device Checker Catches, and What It Does Not

Proofreading Without Uploading: What an On-Device Checker Catches, and What It Does Not

T
Toolz Team
|Nov 2, 2026|12 min read
Prefer Toolz.dev on Google

Part of the Text Tools collection

Grammar Checker

Check grammar, spelling and punctuation in your browser with Harper. Your text is never uploaded, and it keeps working offline.

Use the Grammar Checker

I gave a checker this sentence: "She defiantly agreed to the terms, and the affect was immediate." It flagged affect and told me to use effect. It said nothing at all about defiantly, which should have been definitely.

Both of those are the same kind of mistake, a word swapped for one that sounds like it. One was caught and one was not, and the difference between them explains most of what you need to know about how these tools work and where to stop trusting them.

TL;DR: A rules engine checks spelling, agreement, punctuation and a fixed list of confused words. It is fast, specific, and runs on your device, so nothing you write is uploaded. It caught 160 issues in a 1,360-word document in 95 milliseconds. It missed defiantly for definitely and precede for proceed, because both are correctly spelled words that only context rules out. Use it for the mechanical pass and read for sense yourself. The grammar checker on this site runs Harper in the browser tab.

What does a rules engine actually catch?

Four classes of problem, reliably. All of the numbers below come from running the engine against text written to break it.

Misspellings. The obvious case, and the one every tool handles. conferance comes back with conference at the top of three suggestions.

Agreement. Subject and verb disagreeing in number. "The team have decided" and "we will joins them" are both flagged, and this is the class that most often survives a read-through, because the sentence sounds fine out loud.

Confused words from a known list. irregardless returns Regardless. should of returns should have. then used where than belongs returns than. These are not inferred from context; they are rules, written down, because the same handful of substitutions account for a large share of the errors in English prose. Merriam-Webster has a whole entry on irregardless and the fact that it has been in use since 1795 is not a defence of it in a business email.

Punctuation and spacing. Doubled spaces, missing apostrophes, the comma splice. Low-value individually and worth having, because they are the ones a reader notices without being able to say why.

Here is a document that breaks all four, with what came back:

Input Kind Suggested
Their going to review it Grammar They're
conferance Spelling conference
Irregardless Spelling Regardless
we should of booked Word Choice should have
earlier then we did Miscellaneous than
the affect was immediate Word Choice effect

That document was 1,360 words. It returned 160 issues in 95 milliseconds, after a one-off engine startup of under half a second. Speed is the quiet advantage of a rules engine: there is no request, no queue, and no model to wait for, so the result appears while you are still typing the next sentence.

What does it miss?

Two sentences from the same test, both of which it passed clean:

"She defiantly agreed to the terms." Defiantly is a real word, correctly spelled, in a grammatically valid position. Only the meaning of the sentence rules it out, and a rules engine does not have the meaning. This is the single most common error class in writing produced by people typing quickly, and it is the one you cannot delegate.

"We will precede with the plan." Same shape. Precede and proceed are both words; only context decides.

"The capital of France is Berlin." Flagged nothing, correctly. It is not a grammar error. It is just false. No proofreading tool of any kind is responsible for that, and it is worth saying out loud because people do assume a checker that catches affect will catch a wrong date or a wrong name.

There is a fourth category, which is more interesting: errors it notices and then advises badly. Given "He past the test", it flagged the verb form and suggested pasts and pastes. The right answer is passed. It found the problem and had nothing useful to say about it, which is a failure mode worth recognising, because a suggestion button is very easy to click without reading.

What you wrote What you meant Caught?
the affect was immediate effect Yes, with the right fix
He past the test passed Flagged, wrong suggestions
She defiantly agreed definitely No
We will precede proceed No
The capital of France is Berlin Paris No, and not its job

Why does the dialect setting change the result so much?

More than people expect. I ran one seven-word sentence, "We organise the colour catalogue by centre", through the same engine twice:

  • American English: four errors. organise, colour, catalogue and centre all flagged as misspellings, with organize, color, catalog and center suggested.
  • British English: zero errors.

Same sentence, same engine, same second. A checker fixed to one dialect is simply wrong for everyone writing in the other, and the default in most tools is American. If you are British, Australian, Canadian or Indian and a checker has ever told you that colour is a typo, that is the entire explanation, and the fix is a dropdown rather than a different tool.

This also matters in the other direction. If you write British English for a US audience, setting the dialect to American turns the checker into a localisation pass that catches every spelling your reader would find odd.

What is the practical difference between on-device and online?

Where the text goes. An online checker sends what you wrote to a server that belongs to a company, and whatever the policy says, the text has left your machine and now exists somewhere you do not control. An on-device checker compiles the engine to WebAssembly and runs it inside the browser tab, so there is no request to send.

The reason that is worth caring about is not abstract. The writing people most want checked is often the writing they least want stored: a resignation letter, a legal complaint, a medical summary, a performance review about a named colleague, a client's unpublished draft. Several of those are things an employer or a client has explicitly told people not to paste into third-party tools, which is why plenty of good writing gets sent out unchecked.

You can verify the claim rather than trusting it. Open the page, let it load, switch off your network, and keep typing. The checking carries on, because the engine is already in the tab. A tool that needed a server would stop at the first keystroke.

The trade is the download. The engine is about 19 MB of WebAssembly, which includes the dictionary and the rules, and it arrives once per visit before the first check. On a fast connection that is a couple of seconds; on a slow one it is noticeable. That is the honest cost of not sending your text anywhere, and it is a one-off rather than a per-check cost.

Rules engine, on device Language model, on a server
Where your text goes Nowhere To a company that keeps it
Speed 95ms for 1,360 words A network round trip, per check
Catches mechanical errors Yes Yes
Catches context errors like defiantly No Often
Rewrites a paragraph No Yes
Explains which rule fired Yes Not reliably
Works offline Yes No
Cost None Usually a subscription for the useful half

Neither column is the right answer for everything. A rules engine is the better mechanical pass and the only safe option for text that must not leave your machine. A language model is better at the things that need understanding, and you accept the upload as the price.

How should this fit into an actual edit?

The order matters more than the tools.

Write first, check last. Running a checker while drafting interrupts the only part of the job that needs continuity. The errors will still be there in twenty minutes.

Read the errors before the suggestions. The error says what the engine thinks is wrong. The suggestion is its guess at a fix, and as past to pasts shows, those are not the same quality. Clicking suggestions down a list without reading them is how a clean document acquires a new mistake.

Do the sense pass yourself, after. This is the pass that catches defiantly, the wrong name, the paragraph that contradicts the one above it, and the sentence that says the opposite of what you meant. Nothing described in this guide does it for you. Reading your own text aloud remains the cheapest way to find them, because your mouth stumbles where your eye skips.

Check the numbers separately. Word counts, character limits and reading time are a different question from correctness. The word counter handles the first two, and if you are writing to a hard character limit for a meta description or a post, the character counter is the one you want open beside the draft.

Compare drafts when you have edited heavily. After a long revision it is genuinely difficult to see what changed. The text diff checker puts the two versions side by side, which is also how you catch the paragraph you meant to put back and did not.

Does this replace a proofreader?

No, and the distinction is sharper than the marketing around these tools suggests.

A checker finds errors in sentences. A proofreader finds errors in a document: the heading level that skipped from two to four, the figure referred to as "above" that is now below, the client's company name spelled two ways on facing pages, the sentence that is technically correct and will be read the wrong way by exactly the person it is addressed to. Those are not harder versions of the same task, they are a different task.

What a good mechanical pass does is make the human pass worth more. A reader who is tripping over their and they're on every page never gets to the argument. Clearing the mechanical layer first is what lets someone read for sense, whether that someone is a colleague, an editor, or you tomorrow morning.

Frequently asked questions

Comments

0 comments

0/2000 characters

No comments yet. Be the first to share your thoughts!