Ann’s at it again!
For those who are not familiar with this new age of AI, it is coming at us in a variety of ways. Not only do you have the option to choose which AI system you want to use at any given moment - Gemini, Chat GPT, or in our case Claude (and, of course, nothing stops you from using multiple services) - several now come with a variety of . . . let’s call them flavors or tools given I’m not that much of a tech-head. I’m sure Ann and Devon would be glad to fill in the details using the right terminology; I’m not a Luddite, but my knowledge and vocabulary are minimal. For now I’ll stick with Claude in my discussion, because that’s what I know best. Claude really has three different tools, and several different engines that can run each tool (each burning more tokens - read: $ - for doing what it does). Think, do you really want to drive your 12-cylinder, gas guzzling Ferrari to the grocery store to pick up that carton of eggs you forgot to get yesterday, or would you rather drive your 70mpg hybrid? Some things don’t need a supercar to get the job done.
The first tool is the plain old Claude Chat. It’s a chat-based AI where you ask it a question and it gives you an answer, perhaps asking you to clarify something or ask if you want more detail than what was given. You use it when you want to find out a certain tidbit of information or to think your way through something. Some of my recent Claude chats include: Can Starlink do wifi calling if you’re outside of cellphone coverage? What was the size of a French division in WWI? What is the U.S. President’s annual salary? What his the COVID-19 vs flu incubation periods? What is the classification and taxonomy of a Meerkat? What is the production cycle for a tomahawk missile - how long will it take for one to be delivered if a country ordered one today? Basically, things I used to google to find out answers to (or go to Wikipedia), I just go to Claude Chat for. The nice things about Claude Chat is that the responses are usually a bit more in-depth than a quick google response and it’s pretty fast.
The next step up in the world is Claude Cowork. Claude Cowork is designed to do more complex search tasks with the end goal of producing some sort of product. It often involves a multi-stage, iterative process of explaining what you want, Claude coming back with something that’s not quite right (or coming back with a series of questions, before coming back with something not quite right), you adjusting things a bit and Claude doing it again until you get an output that is satisfactory. It’s intended to produce something as an end product.
For example, I could say that I want to create a comprehensive guide to each of the 48 teams that are participating in this year’s World Cup. Each team has submitted its initial roster of up to 55 players, from which 26 will be selected just prior to the start of the World Cup. I would like you (CLAUDE) to build me an excel spreadsheet of each of the teams and each team’s initial roster, also identify each of the players who are likely to make the 26-man roster, who might be injured and not make the roster, and who are the 11 players likely to be starters for each team, and if some players are in close competition, flag them as well. Also, prepare me a Word document summary for each team - their coach’s footballing philosophy, their offensive and defensive strengths and weaknesses, and who their star players are and who their team leaders are. Oh yeah, and I want you to summarize the team information into a power point presentation that discusses the likelihood of each team winning the World Cup and why, starting with the favorites and ending with the long-shots. I suspect if I did that, I’d have an Excel spreadsheet, Word team summaries and a power point in less than an hour [now that I’ve thought about it, I may try it . . . ok, so I did, though I was dumb about it, giving it everything at once used up all of my session time, so it timed out on me, twice. Though my guess was pretty much right - I timed two 15 minute sessions of intense research and compilation, and it was about half done so . . . an hour. I decided not to finish the project.].
The nice thing about Claude Cowork is it’s designed for folks like me. People who can formulate in their mind what they want and an outcome product, but who have absolutely no idea about programming. Claude Cowork, if necessary, will create programs to do the tasks you’ve asked it to do. I’m still trying to figure out how to maximize it in my own work (I’m proceeding very cautiously due to malpractice issues), but it is the future of work.
Then there is Claude Code. That’s where the tech heads (read: Ann) can really go wild. For folks who know programming, they can push Claude to create programs that do some very interesting things, and they no longer need to know the “language” necessary to do that type of programming. I’m lost in that realm because, from the few videos I’ve watched about using it effectively, it works best when operating from a “terminal window” (no, I don’t really know what that is other than I’ve seen them open on Ann’s computer for years) and operates using “markup language” (I kinda know what that is, but I don’t speak markup). But Ann does . . . both.
So a couple of weeks ago I made one of my stops downstairs taking a break from brief writing. “So, what have you been working on?” Ann replied, “I just finished up a Lightroom plug-in!” “Wait . . . what?” “Yeah, remember what I was working on before, getting Claude to help me with rating images and stuff? Well, I figured why not try to turn that into a Lightroom plug-in so I can use it straight from Lightroom!” Unlike the previous sessions, which used Claude Chat or Cowork, this time she used Claude Code to create the plug-in. Like I said - tech heads.
I asked for a demo. Ann pulls up one of the photographs she took while in Sweden.
She goes up to the menu bar in Lightroom, selects “Plug-in Extras” and “Analyze Selected Photos with AI” and . . .
And it does its thing, putting the analysis into the metadata for the photograph.
So what does the plug-in do? [1] it identifies a category it believes the image belongs (Ann and Claude have decided that for cataloging and website purposes, an image should belong to only one category); [2] If it’s a close call between categories, the competing category will be placed in the “AI Boundary” section; [3] then there is a recommended Star rating - only images 3 stars or higher get rated; [4] then there is a recommendation for whether the image should go online into a Web collection, with possible responses being Yes-No-Maybe; [5] Then there is an estimated AI Location - Ann has found this to be truly a wild guess, it’s often quite wrong, but it’s interesting to see why it thinks what it does; [6] next is a listing of recommended keywords - Ann of course has another plug in (purchased from another person) that easily and quickly moves the AI recommended keywords Ann wants to use into the Keyword metadata panel in Lightroom; [7] then comes the key analytical sections, first with an AI Critique; [8] followed by the AI Strengths; [9] and then the AI Concerns about the image. [10] Last is a date file the AI analysis was created. The result looks like this under the Metadata tab for the image in Lightroom:
And it takes about 10-20 seconds. Not instantaneous, and it takes longer when you do a group of images, but it’s surprisingly fast.
As I mentioned before, Ann doesn’t always take the recommendations - she’ll often disagree with the star rating (up or down), and with some of the analysis (more on that in another post), but it’s been a very useful tool. One afternoon, while reading the AI Concerns for one of the images Ann had, I said, “Well, I think we could fix that with a bit of development.” And we did, which made for a stronger image.
So kudos to Ann for her plug-in. I certainly never thought of that!