I use Grammarly to correct typos, grammar mistakes, and sometimes some style issues. I’ve been using Grammarly for years, as I have a paid account through work.

Since Grammarly doesn’t work on Zen on my Linux desktop for some reason (it refuses to sign into my account, and I suspect it’s my VPN I have there, which I’m not going to turn off), I was looking for a solution. There is the pretty good LanguageTool, but I got curious about using AI to do more than just a quick grammar check.

With folks around me using AI for creating more sophisticated tools, especially after reviewing Doctorow’s usage of AI and his viewpoint on using such tools, I thought of using it to help me with some of the editing itself. Since I use Kagi, I could utilize its custom assistant feature, which allows me to choose an LLM from an available list and then build and save a specific prompt to use each time I want to work on a post.

There’s a lot that goes into long-form1 blog posts. I believe I went into that in detail before, but here is my process again, especially since I started to be more “hands on” in different social media platforms:

  1. Write the draft in Emacs org-mode
  2. Export to markdown inside Emacs, copy-paste into the Micro.blog macOS app
  3. Correct grammar and style in Grammarly inside the Micro.blog app
  4. Select a category, add a summary, and decide which platforms I’m going to cross-post the post to automatically (these are usually Mastodon and Tumblr, which allow me to edit posts; Blue Sky doesn’t allow that, which means I need to create another post there manually)
  5. Go into the different social media platforms and polish those up (character limit issue, adding hashtags, etc.)

I was blown away by how well this worked from the get-go.

Here’s what I’m asking Kimi, the AI model I chose for that task, to do

  1. Explain who I am (the author of this blog, what pronouns to use, how to call me)

  2. Scan this blog, and look for posts that “greatly overlap” the draft I just submitted. Kimi has been smart enough so far to figure out when I’m basically rambling about something I already did, or if it’s close but I have something new to add (or if it’s new content altogether)

  3. If it finds an identical post, it stops and asks me if I want to proceed anyway.

  4. If I say I want to continue, it starts looking for spelling mistakes and grammatical errors. It corrects me automatically, but in a separate section under the draft, it tells me what it corrected. This is useful because it’s not always correct, and I want to make sure it doesn’t kill my awesome grumpy style.

  5. It goes over the links of my post to check if any of them are problematic. I was delighted to see it caught a mistake where I pointed two different links to the same URL, a simple copy-paste error, and brought it to my attention.

  6. It recommends terms I should expand on, while offering references. I asked it to prioritize official links and Wikipedia links, and it works well. Additionally, it will scan this blog and let me know if any of my other posts already explain the terms in my current draft, summarize them, and provide links to those posts.

  7. Finally, it gives me a list of hashtags that it thinks I should use for Mastodon, Blue Sky, and Tumblr. This part is the weakest, as these trends change frequently and don’t always align with what my post is about, but it’s a good starting point. I edit those manually as needed.

The key to using AI, in my opinion, is to always see what it finds and inspect its reasoning if something feels off. Because I’m using a reasoning model, I can see what it was “thinking” and where something went wrong; this is how I’ve been able to polish it over the last few days.

I’d love it if I could get Kimi to also scan my old blog at https://master--taonaw-blog.netlify.app, but netlify.app isn’t indexed by AI bots or Google (which Kagi is built on mostly), so it doesn’t work. I could probably work around this if I upload a text of my old blog or post it somewhere else, if I get to it.

Here’s a glimpse at how it looks like when I let it work through this very post:

Footnotes

1: In general, I split my blog posts into quick “bursts,” sort of “tweets” or “toots” which are quick and easy, and then there are the longer posts which take more time. These tend to get edited more deeply, as I like to use different sources and include visual aids, among other things. Those longer posts are the kind of posts I’m discussing here.