The Grammar feature seems to have weird suggestions/cycles too on a little bit more testing. Curious to see how this improves. A local only, one-time-purchase grammarly alternative is appealing!
[1] Loosely; I’d say not referencing any networking entrypoints or dlsym() also counts, as working around that would be very non-deniably malicious.
Edit: Alas, Hacker News also removes the extra space after periods.
Does not catch a singular/plural discrepancy between the subject and the verb in a sentence--a common mistake when the expressed thought applies equally to one thing or to many things.
Does not catch a missed indefinite article--a common mistake for speakers of languages that don't have articles. Similarly, does not catch the use of the indefinite article for a thing already mentioned before.
Does not even catch the obvious "don;t" typo.
The fluency suggestions are seemingly largely malfunctioning. It frequently suggests starting and ending sentences with quotes, although it also makes some useful suggestions. There seems to be an issue with analysis running synchronized, or something like that; when I type into a text field and Refine starts to run, it often blocks text entry. Selecting a suggested replacement blocks the app for half a second or so. Neither of these problems occurs with Grammarly or Language Tool. I also noticed a bunch of issues that Grammarly catches (like verb agreement) that Refine does not.
But this is an amazing first release and extremely promising. Congrats!
I realize I'm a niche :)
Does anybody know of such a tool?
Two of your 4 questions were answered in the first content block
This phrase is offensive and violates my safety guidelines. Therefore, I will not revise it. I am programmed to avoid generating responses that are obscene, or that contain profanity.
unsloth/gemma-3n-E4B-it-GGUF
You are a precision editor guided by a custom style manual. Your tasks are ordered by priority.
Your primary rule is to consult the provided REFERENCE DICTIONARY. Any term on this list is correct and must be preserved exactly as written.
Your secondary rule is to refine phrasing and sentence structure to improve clarity, conciseness, and flow. The goal is to make the text read more naturally and professionally, while **strictly preserving the author's core meaning and tone.**
Your final rule is to output ONLY the clean, revised text. You MUST NOT add any commentary, greetings, or explanations.
REFERENCE DICTIONARY:
{{dictionary_words}}
Revise the following:
"{{sentence}}"
You are an expert editor. Your single most important goal is to improve the fluency and clarity of the following text while STRICTLY PRESERVING the author's original voice and meaning.
You MUST follow these rules:
1. Only rephrase sentences that are genuinely awkward or unclear.
2. Never make changes for purely stylistic preference.
Return ONLY the clean, revised text.
Revise the following:
"{{sentence}}"
{{dictionary_words}}
You are a silent grammar correction engine with a custom style guide.
Your primary rule is to consult the provided REFERENCE DICTIONARY. Any term on this list is correct and must be preserved exactly as written.
Your secondary rule is to correct all other grammar, spelling, and punctuation errors in the main text.
Your final rule is to output ONLY the clean, corrected text. You MUST NOT add any commentary, greetings, or explanations.
REFERENCE DICTIONARY:
{{dictionary_words}}
Correct the following:
"{{sentence}}"
You are a silent grammar correction engine. Your sole function is to receive text and output the corrected version. You MUST NOT add any commentary, greetings, or explanations. You will only return the clean, corrected text.
Correct the following:
"{{sentence}}"> Sometimes I still make mistakes with articles and prepositions, but my grammar is getting better every day I practice.
In American/Simplified English, this is grammatically correct. However, in 'full fat' English, practice is a noun, whereas practise is a verb; e.g.:
> I go to my practice to practise medicine.
The problem I have with this website is that it's entirely concerned with peripheral issues. The product respects my privacy - good. The product is performant - good. The product doesn't require an Internet connection - good. The product works in many writing apps - good. The product has transparent pricing - good. But I don't give a shit about any of this until you convince me that this will consistently do the correct thing, and this website singularly fails to achieve this.
> ... in 'full fat' English ...
English is a bastard of a language and getting messier every day as new nations adopt it is their standard language.
Setting the bar where it is well written and unambiguously understandable is IMHO completely fine for a 15$ product.
Having a text spell checked to comply with contemporary Oxford English is likely not the goal of this product.
I disagree strenuously with this idea, because it suggests that there is one 'big' English in which anything goes. A better idea is the one of the register[0]: there are many Englishes, many sets of rules. Different rules are used in different regions, by different groups of people, and have different connotations (e.g., the King James Bible was intentionally written in a form of English that was considered archaic at the time because that would make it sound more grandiose).
If I were to use this tool, I'd be using it to ensure that whatever I'm writing is well-received by my intended audience. Because English usage is so varied, I would need to be able to control the register that it uses to ensure that the output is suitable. The fact that the product website doesn't even mention a list of supported languages, let alone supported dialects and registers within those languages, has a very everyone can see what a horse is kind of feeling[1].
[0] https://en.wikipedia.org/wiki/Register_(sociolinguistics)
What you suggest does seem like a good early doors feature; but the cut they've made seems to be the right one to prove market potential.
I worry that this will make my writing more likely to fail an AI coursework detector, which could really impact my life. The risk just isn't worth it till someone has tested the output through all the big players (turnitin etc.)
Colour and licence are so quaint.
We have enough American cultural hegemony as it is. It frustrates me no end that I regularly am unable[1] to configure software to use my preferred version of English.
[1] or it's extremely difficult. CLion I'm looking at you; every available option I have is set to British English, but still you insist on telling me my Colours should be corrected to Color etc. :(
And think global. A lot of private schools around the world require students to use Macs so no software update. Have a British English option for outside North/South America as, aside from some edge cases, British English is more common at a school level.
gbalduzzi•3h ago
runjuu•3h ago
In upcoming releases, Refine will support custom prompts and BYOK (Bring Your Own Key), allowing you to use any large language model you want.
pyman•1h ago
You do realise you're already using an optimised model built for everyday devices, and that model includes some serious innovations in parameter-efficient processing, right?
You're a great developer, and it looks like you're thinking about adding features like BYOK quickly to please more users. But in doing that, you might be missing the real innovation you've already created. You've basically built a version of Grammarly without the privacy issues that make most IT departments ban it.
No one wants Grammarly or your tool sending corporate emails or documents to a language model.
Look at Apple. Privacy is what sets them apart. And right now, privacy is your biggest feature.
Use that. Add a big table to your site comparing your products privacy with Grammarly's. That's your strongest selling point, and probably the only feature that can truly compete with the big players.
My advice? Keep improving the app, but keep the model local. Keep it private. That's the killer feature you've got.