A very pragmatic guide addressing many points that are not evident when heading in this direction. I can see parallele with my own evolving approach but unfortunately think that the clarity you bring won't be realised without experiencing a few knocks along the way. Well said, Alex.
Interesting piece and quite enjoyed it. I especially like the idea of auto-translated websites. I used to have a github pages but recently replaced with a Carrd landing page. But now I'm re-thinking. Many thanks.
A few follow up questions come to mind about other parts of the piece.
1. You start with a style guide and let it evolve after you make edits. What's a basic one look like? I get why you don't share yours, because then that would color everyone else's. I presume I could start with something like a crib sheet from Chicago, MLA, or whatever, add rules from Strunk & White, and then include any unique things for me (say, "always 'policy-maker' not 'policymaker'" or the other way around depending on each person's style guide). But how long should that first one be, and then how do I decide which edits need to be moved back into the styleguide and which are just one-offs that can be left alone? What else is missing here?
2. I have github and use it sporadically. How does the agent work with github? Is it reading the repo or is it reading the version control from the `.git`files in a folder?
3. Speaking of agents. I'm not sure I follow. Is this the same thing as having the desktop version of the AI or something that each scholar builds for himself?
4. Relatedly, what's your recommendation for scholars who use university issued computers that prohibit agents on their machines? My university laptop is extremely locked down. All software has to be approved through a workday ticket. Many requests are often denied for "security" reasons (read: the university IT dept doesn't want to deal with it at scale). As a result, I have a workflow that relies on my MacBook Neo for all writing and custom software, I use my work issued MacBook Pro for quant work. It's cumbersome but works. I suspect others might only have their work laptop and cannot use an agent or desktop versions.
Thanks, Luke. I may need to do a follow-up piece on some of these points, but very briefly:
1) I would start by putting together all your writing, converting it to .md, and asking your agent to do an analysis of it explicitly for the purposes of your style guide. It should be able to pick up things like "policy-maker." Then I would ask it to generate a few short documents based on your ramblings, edit them yourself, and ask it to analyze the changes you have made.
2) The agent works in a regular local folder that is also a git repository, so it reads your files directly and uses the version history to undo or compare things when needed. GitHub is just the cloud copy of that folder. It's not strictly necessary, but I recommend it for backup and for free website hosting, and a private repo is fine for anything you don't want public.
3) It's an off-the-shelf tool you install and log into (Claude Code, Codex), not something each scholar builds for himself, and it's separate from the desktop chat apps. I would recommend the resource I linked earlier: https://www.oneusefulthing.org/p/an-opinionated-guide-to-which-ai-b22 . But basically, the chatbot tells you stuff, while the agent does stuff (which can happen in the cloud/GitHub too, even without any desktop access).
4) I would recommend having a decent home laptop/computer for exactly this reason. For work that has to stay on the locked-down machine, the browser-based cloud versions of these agents can sometimes get around the install problem, since nothing runs locally, but that depends on your IT policy too.
Thanks for the pro-set up. I have started on this in a looser way, but will definitely try. Funnily enough, I also just did a post why academics are so rubbish at writing for a general audience (curse of knowledge, information gap theory) and came across an old Pinker piece which was a brilliant read.
So I am totally in favour of academics using tools to get them over the gap. Sadly, I then I discovered that some of my colleagues, including those much younger than me, use AI but are not capable of editing their text to make it come back as not majority AI generated. Bit gobsmacked at that. (Trick: don’t change words, change syntax!)
No idea what will happen with the new watermarking… Do you have a workflow for that yet? :-)
Huh. I didn't know it was a syntax vs word distinction. I likely have been doing this intuitively but it makes sense. AI is like having a student or co-author write a very rough first draft but too many of us academics still default to a simple copy/paste interaction, which baffles me.
Really useful, Alex. Not so much for me to adopt an agent in full (thats not where my bottleneck is) but to get me think seriously abt how I use AI. Having a fun exchange w Claude on this. I'll keep you posted :)
Apart from creating specific projects to stress test arguments, two things I'll use Claude in a way I haven't yet - career planning (not the politics, but strategically. Not much mentoring around for mid-career scholars) and for chucking half formed ideas at it to see viability, feasibility, and general worth (so the ideas I get whilst running for a paper, a project, a proposal or whatever). None of these require agents which, if I understand ot correctly, help in the delegation of work. I find Claude most useful as co-thinker and for augmentation. Once I feel more knowledgeable abt it, I want to see how I can incorporate it in my teaching, ie how do I get students to make a constructive use of what I tell them in class and of what they read? How do I keep fostering critical thinking in the 21st century?
The weekly translation drift check is the part I'd keep. Creating eleven versions is a one-time job. Catching the moment they quietly fall out of sync is the operating problem.
A very pragmatic guide addressing many points that are not evident when heading in this direction. I can see parallele with my own evolving approach but unfortunately think that the clarity you bring won't be realised without experiencing a few knocks along the way. Well said, Alex.
Thanks, Michael!
Interesting piece and quite enjoyed it. I especially like the idea of auto-translated websites. I used to have a github pages but recently replaced with a Carrd landing page. But now I'm re-thinking. Many thanks.
A few follow up questions come to mind about other parts of the piece.
1. You start with a style guide and let it evolve after you make edits. What's a basic one look like? I get why you don't share yours, because then that would color everyone else's. I presume I could start with something like a crib sheet from Chicago, MLA, or whatever, add rules from Strunk & White, and then include any unique things for me (say, "always 'policy-maker' not 'policymaker'" or the other way around depending on each person's style guide). But how long should that first one be, and then how do I decide which edits need to be moved back into the styleguide and which are just one-offs that can be left alone? What else is missing here?
2. I have github and use it sporadically. How does the agent work with github? Is it reading the repo or is it reading the version control from the `.git`files in a folder?
3. Speaking of agents. I'm not sure I follow. Is this the same thing as having the desktop version of the AI or something that each scholar builds for himself?
4. Relatedly, what's your recommendation for scholars who use university issued computers that prohibit agents on their machines? My university laptop is extremely locked down. All software has to be approved through a workday ticket. Many requests are often denied for "security" reasons (read: the university IT dept doesn't want to deal with it at scale). As a result, I have a workflow that relies on my MacBook Neo for all writing and custom software, I use my work issued MacBook Pro for quant work. It's cumbersome but works. I suspect others might only have their work laptop and cannot use an agent or desktop versions.
Thanks, Luke. I may need to do a follow-up piece on some of these points, but very briefly:
1) I would start by putting together all your writing, converting it to .md, and asking your agent to do an analysis of it explicitly for the purposes of your style guide. It should be able to pick up things like "policy-maker." Then I would ask it to generate a few short documents based on your ramblings, edit them yourself, and ask it to analyze the changes you have made.
2) The agent works in a regular local folder that is also a git repository, so it reads your files directly and uses the version history to undo or compare things when needed. GitHub is just the cloud copy of that folder. It's not strictly necessary, but I recommend it for backup and for free website hosting, and a private repo is fine for anything you don't want public.
3) It's an off-the-shelf tool you install and log into (Claude Code, Codex), not something each scholar builds for himself, and it's separate from the desktop chat apps. I would recommend the resource I linked earlier: https://www.oneusefulthing.org/p/an-opinionated-guide-to-which-ai-b22 . But basically, the chatbot tells you stuff, while the agent does stuff (which can happen in the cloud/GitHub too, even without any desktop access).
4) I would recommend having a decent home laptop/computer for exactly this reason. For work that has to stay on the locked-down machine, the browser-based cloud versions of these agents can sometimes get around the install problem, since nothing runs locally, but that depends on your IT policy too.
Thanks for the pro-set up. I have started on this in a looser way, but will definitely try. Funnily enough, I also just did a post why academics are so rubbish at writing for a general audience (curse of knowledge, information gap theory) and came across an old Pinker piece which was a brilliant read.
So I am totally in favour of academics using tools to get them over the gap. Sadly, I then I discovered that some of my colleagues, including those much younger than me, use AI but are not capable of editing their text to make it come back as not majority AI generated. Bit gobsmacked at that. (Trick: don’t change words, change syntax!)
No idea what will happen with the new watermarking… Do you have a workflow for that yet? :-)
Huh. I didn't know it was a syntax vs word distinction. I likely have been doing this intuitively but it makes sense. AI is like having a student or co-author write a very rough first draft but too many of us academics still default to a simple copy/paste interaction, which baffles me.
Really useful, Alex. Not so much for me to adopt an agent in full (thats not where my bottleneck is) but to get me think seriously abt how I use AI. Having a fun exchange w Claude on this. I'll keep you posted :)
Thanks, Sarah! I'm quite curious where something related to my approach may lead folks :)
Apart from creating specific projects to stress test arguments, two things I'll use Claude in a way I haven't yet - career planning (not the politics, but strategically. Not much mentoring around for mid-career scholars) and for chucking half formed ideas at it to see viability, feasibility, and general worth (so the ideas I get whilst running for a paper, a project, a proposal or whatever). None of these require agents which, if I understand ot correctly, help in the delegation of work. I find Claude most useful as co-thinker and for augmentation. Once I feel more knowledgeable abt it, I want to see how I can incorporate it in my teaching, ie how do I get students to make a constructive use of what I tell them in class and of what they read? How do I keep fostering critical thinking in the 21st century?
The weekly translation drift check is the part I'd keep. Creating eleven versions is a one-time job. Catching the moment they quietly fall out of sync is the operating problem.