Does anyone have a checklist for things to do before publishing a blog post? “check for unexplained or unnecessary jargon & acronyms” “make sure there’s a summary at the top”, “try to put in more diagrams / images” … that kind of stuff. (Kinda embarrassing that I haven’t put together my own long ago, but better late than never I guess.)
Relatedly, when I write a blog post short enough to fit the whole thing in the GPT4 context window, then I can ask GPT4 to list out grammar errors, jargon, etc. If anyone has also been doing that, I’d be interested to hear what prompts you use. Thanks in advance.
- [ ] Check for unexplained or unnecessary jargon & acronyms. - [ ] Check for jargon & acronyms that are defined in one part of the post and then used in a distant part of the post without repeating the definition. - [ ] Check for unnecessarily obscure words and cultural references (for non-native English speakers) - [ ] Check for vague “this” - [ ] Check for over-hedging - [ ] Check all hyperlinks - [ ] Look for places to add hyperlinks, references, and footnotes - [ ] Consider adding a summary / ToC / tldr to the top - [ ] Consider adding funny things especially near the top - [ ] Look for places to add pictures, possibly including DALL-E 2 etc. - [ ] Look for places to add concrete examples - [ ] Make sure all pictures have a white background, not transparent (for dark-mode readers). (Check by just viewing it in dark mode.) - [ ] Check the LW sidebar outline looks right
GPT copyediting prompts
(I should experiment more before sharing)
Consider sharing the draft with people
- [ ] Consider friends like [long redacted list] - [ ] Consider Slacks/discords/etc like [long redacted list] - [ ] Consider anyone whose paper I’m citing - [ ] Consider anyone who I mention by name - [ ] Consider anyone deeply involved in a field that I’m talking about
Consider sending it to LW copyediting (Justis)
(Always use the “ask for feedback” button, not the general intercom, so that it goes straight to Justis rather than a Lightcone staff person needing to forward it. In particular, if it’s currently in gdocs rather than LW, then just put the title and the gdocs link into an otherwise-blank LW editor and press the “ask for feedback” button.)
Repeat the copyediting list from above one more time if there have been changes
After publishing
- [ ] Tweet & Mastodon - [ ] Add to my webpage
~~
Thanks @Max H for some of those. It looks intimidating but I already basically do all these things (or explicitly consider doing it and decide not to), the only problem is I sometimes forget a few of them. This should help!
Can you proofread the first few sections of my draft post for lesswrong for spelling / grammar, and suggest any improvements to style, clarity, and brevity? Please provide a corrected draft, followed by a list of all the changes you made. Also please provide a summary of the entire post. The draft should be formatted in Markdown, enclosed in a code formatting block.
Followed by pasting in parts of my post, leaving enough room in the context window for GPT-4 to respond.
It works OK, but I don’t always like the style suggestions it makes. I use vim and git diffs to edit my drafts and figure out which edits from GPT-4 I actually want to accept.
Some checks I do manually when proofreading my own posts:
check for words or phrases that are repeated too closely together
check for vague “its” or “this” (which is kind of the opposite of repeating words too often)
check all hyperlinks
check for anything that should be hyperlinked / ref’d that isn’t
check formatting of code blocks, titles, and other md or LW docs elements
check the sidebar outline looks right
read the whole post backwards, sentence-wise.
I also found Justis’s editing guide helpful, as well as his actual editing and proof-reading service (available by clicking “get feedback” on a draft post).
I asked GPT-4 to improve on the handwritten prompt above. After a couple of iterations, it came up with this:
Proofread the first few sections of my draft post for LessWrong, focusing on spelling, grammar, punctuation, repeated words/phrases, sentence structure, and the clarity of pronouns such as "this" and "it". Additionally, please suggest improvements to style, clarity, brevity, logical flow, and coherence of arguments. Provide a corrected draft in Markdown format, enclosed in a code formatting block, followed by a list of all the changes you made. Lastly, please provide a summary of the entire post.
Which I might try next time. There are lots of folklore tips for improving performance by telling the model it is an expert or whatever, so there might be a lot more room for improvement here.
Does anyone have a checklist for things to do before publishing a blog post? “check for unexplained or unnecessary jargon & acronyms” “make sure there’s a summary at the top”, “try to put in more diagrams / images” … that kind of stuff. (Kinda embarrassing that I haven’t put together my own long ago, but better late than never I guess.)
Relatedly, when I write a blog post short enough to fit the whole thing in the GPT4 context window, then I can ask GPT4 to list out grammar errors, jargon, etc. If anyone has also been doing that, I’d be interested to hear what prompts you use. Thanks in advance.
Current draft is:
Copyediting
- [ ] Check for unexplained or unnecessary jargon & acronyms.
- [ ] Check for jargon & acronyms that are defined in one part of the post and then used in a distant part of the post without repeating the definition.
- [ ] Check for unnecessarily obscure words and cultural references (for non-native English speakers)
- [ ] Check for vague “this”
- [ ] Check for over-hedging
- [ ] Check all hyperlinks
- [ ] Look for places to add hyperlinks, references, and footnotes
- [ ] Consider adding a summary / ToC / tldr to the top
- [ ] Consider adding funny things especially near the top
- [ ] Look for places to add pictures, possibly including DALL-E 2 etc.
- [ ] Look for places to add concrete examples
- [ ] Make sure all pictures have a white background, not transparent (for dark-mode readers). (Check by just viewing it in dark mode.)
- [ ] Check the LW sidebar outline looks right
GPT copyediting prompts
(I should experiment more before sharing)
Consider sharing the draft with people
- [ ] Consider friends like [long redacted list]
- [ ] Consider Slacks/discords/etc like [long redacted list]
- [ ] Consider anyone whose paper I’m citing
- [ ] Consider anyone who I mention by name
- [ ] Consider anyone deeply involved in a field that I’m talking about
Consider sending it to LW copyediting (Justis)
(Always use the “ask for feedback” button, not the general intercom, so that it goes straight to Justis rather than a Lightcone staff person needing to forward it. In particular, if it’s currently in gdocs rather than LW, then just put the title and the gdocs link into an otherwise-blank LW editor and press the “ask for feedback” button.)
Repeat the copyediting list from above one more time if there have been changes
After publishing
- [ ] Tweet & Mastodon
- [ ] Add to my webpage
~~
Thanks @Max H for some of those. It looks intimidating but I already basically do all these things (or explicitly consider doing it and decide not to), the only problem is I sometimes forget a few of them. This should help!
Still very very open to suggestions. :)
For GPT-4, I’ve used this:
Followed by pasting in parts of my post, leaving enough room in the context window for GPT-4 to respond.
It works OK, but I don’t always like the style suggestions it makes. I use vim and git diffs to edit my drafts and figure out which edits from GPT-4 I actually want to accept.
Some checks I do manually when proofreading my own posts:
check for words or phrases that are repeated too closely together
check for vague “its” or “this” (which is kind of the opposite of repeating words too often)
check all hyperlinks
check for anything that should be hyperlinked / ref’d that isn’t
check formatting of code blocks, titles, and other md or LW docs elements
check the sidebar outline looks right
read the whole post backwards, sentence-wise.
I also found Justis’s editing guide helpful, as well as his actual editing and proof-reading service (available by clicking “get feedback” on a draft post).
I asked GPT-4 to improve on the handwritten prompt above. After a couple of iterations, it came up with this:
Which I might try next time. There are lots of folklore tips for improving performance by telling the model it is an expert or whatever, so there might be a lot more room for improvement here.