Later, when they released their API, I developed a CLI tool for this purpose (https://github.com/pncnmnp/grammatical). Note that it is not flawless, but it works well. It has improved my writing productivity, both for blogging and emails.
I like your prompts for your script, I may integrate it into my workflow too. I have found that when I am editing my work, it is helpful to have the original sentence and the suggested sentence one after another. While this takes away the paragraph form, it clearly helps you tease out the "improvements" to either accept or reject. I also have GPT number the sentences. And then when I am done, I say - something like - "please assemble back into paragraph form, all revised sentences are accepted except numbers 5, 12." And then it reassembles the paragraph(s) with the revised corrections. I use ChatGPT (4) for really long stuff though... might not work with Turbo3.5.
That is an excellent idea! Perhaps someone should consider developing a smart diff tool for this purpose.
To be honest, my work style can be a bit lazy at times, and this may reflect in my approach. My tool simply copies any changes directly to the clipboard. I then replace them with the original text and make adjustments on the same page. However, for more complex workflows, your approach is awesome. It could potentially transform the process into something similar to Grammarly.
Edit: Have you noticed any differences in terms of quality between GPT-4 and Turbo3.5 for this particular task?
Ah, that's interesting! It could be related to the improvements they seem to have made in the area of "overreliance". According to OpenAI's paper (https://arxiv.org/pdf/2303.08774.pdf):
> Overreliance occurs when users excessively trust and depend on the model, potentially leading to unnoticed mistakes and inadequate oversight.
> At the model-level we’ve also made changes to address the risks of both overreliance and underreliance. Weve found that GPT-4 exhibits enhanced steerability which allows it to better infer users intentions without extensive prompt tuning.
> To tackle overreliance, we’ve refined the model’s refusal behavior, making it more stringent in rejecting requests that go against our content policy, while being more open to requests it can safely fulfill. One objective here is to discourage users from disregarding the model’s refusals.
> However, it’s worth noting that GPT-4 still displays a tendency to hedge in its responses.
However, it appears that the self-hosting option only gives access to the basic version (https://news.ycombinator.com/item?id=32238705), which is still impressive, but the premium version supposedly has better grammar and style features.
Honestly, ChatGPT's $0.002 per 1k tokens is quite tempting for me. Even after hundreds of queries, my monthly usage is less than 50 cents.
Later, when they released their API, I developed a CLI tool for this purpose (https://github.com/pncnmnp/grammatical). Note that it is not flawless, but it works well. It has improved my writing productivity, both for blogging and emails.