ClarisTalk AI · the FileMaker AI podcast
Ian Jempson – New tool: Transcribr for FileMaker.
Ian Jempson demonstrates an immediately usefully AI tools for FileMaker developers that transcribes video and audio files. Maybe even bigger: Ian shows his workflow for not only writing script with complex code that can be pasted into Script Workspace (close to the Holy Grail here). Also he shows a way to prompt-into-existence entire database files with tables, layouts, scripts and relationships.
Apple PodcastsSpotifyYouTubeAll episodes
Transcript
Transcribed by machine (whisper, large model), 12,127 words. Names and product terms are sometimes misheard; the audio is the record.
One of the leaders in the FileMaker community that has really been on the forefront of AI is Ian Jemson, who is our guest today on the podcast, Clarisse Talk AI. Ian, welcome, sir. Thanks for having me, Matt. It's good to be back. It's always good. You have been and will continue to be a regular, I hope. So you have some new stuff to announce today. I'm really looking forward to seeing. I've only gotten a little preview hints from you. So you have two tools. One of them you're releasing. Tell me about it. Okay. It's called Transcriber. And what it is, it's a local transcription tool for FileMaker. Version 1 is named Squarely at FileMaker Developers, and it's a completely self-contained application.
First one being released is for the Mac. I have tested the Linux version, and that works. I've not yet built the Windows version. Server-side, yes. Yeah, yeah, yeah, because there's no Linux client. But one of the things I found was I've got a whole load of AI tools that I've been running for up to a couple of years now. And for people to use some of these things, you basically have to go and install a load of dependencies and you need to manage Python versions and worry about a whole load of stuff. And when you install these Python dependencies, You break your Python and your system and all that. So what I've built is it is a single file executable. It's just one single on the Mac.
It's Swift. So it's a Swift app that runs to install by drag and drop. I'm actually about to see if Apple will accept it in the store. There's no reason they shouldn't, but it's signed by Apple. So there's no issues with running it. um anyway so uh would you like to have a bit of a look no we'll just skip it and talk about it of course i would share my screen yeah definitely so the audio listeners i'll be trying to describe stuff uh yeah okay desktop two um okay so that's the safari and then okay here was the installer There's the installer. So basically, it's a disk image. Just agree. Transcriber missing the E on purpose, no doubt. Darn. Yes, that's on purpose.
And literally, you drag it into your applications folder to install. I'm not going to do that because I've got it running at the moment. Yeah, and I see it includes a demo database. Great. Yeah, includes a FileMaker demo database. um if i say we can if running's on localhost it's port 8001 which i chose because it's not normally in use uh that is configurable so if you can copy the url start the service stop oh nice okay setting this window oh yeah yeah okay so a little bit similar to like lm studio sort of like in you you can start and stop except you're not manually loading the model but it's basically kind of running as a AI tool. Absolutely. So version 1 uses the Whisper models.
So-- From AWS, yeah? Whisper is from Meta. Version 1.1, which is about a week away-- yeah, a week away-- that adds in Parakeet from NVIDIA. And that is shockingly fast. it will on my mac m1 uh macbook m1 it will transcribe a file at 90 times real time so it will transcribe an hour-long video or video or audio in about 45 seconds so what about youtube just i'm jumping the gun uh actually yeah there's no reason why it would uh it does the um it does the file conversion on um so uh we basically provide a real-time ui which actually this so you can actually transcribe directly into filemaker oh yes because it is it is kind of a web service but if i just uh click okay to that back and then And then we've got sort of the API documentation.
It's got a full Swagger documentation of everything. So essentially, you interact with it by insert from URL. Yep. So what you can do is you can transcribe a single file. You can batch process a folder. So you can just give it a folder name, do the whole folder at once. Amazing. The other thing you can do is you can watch a folder. So you can say, I want this folder transcribed, everything that gets dropped into this folder, and it will just do that. I'm seeing some kind of immediate use for this. Anyway, plus I see that you have, if you scroll up a little bit, so you've got FileWatcher transcripts, and then you also have a WebSocket. That's correct. So the WebSocket, essentially, that's what allows you to do real-time transcription.
So if I just go to FileMaker, I'll just move that out of the way. So this is your sample file, probably? The sample file looks like this when you launch it. Okay. Basically, it has a list of the endpoints. Okay. A description of how they all work. Nice, clear UI. So that's similar to the Swagger. You've got real-time transcription. This actually probably won't work if I try it here because the microphone setting with Zoom. So real-time transcription, like in FileMaker, you can just kind of tap the microphone button a couple of times and it will just pop up the Apple transcriber where you can just speak. So, I mean, that part is kind of solved. It's the Apple one, which is not as good.
Yeah, I would tend to agree. Yeah, for most purposes, that's what you're going to want to use. But, however, this is 100% local as well. Now, Apple, I think, in the latest OS has probably solved that. I think it's doing as much as it can locally. But it kind of is there, should you want it. You hadn't thought about that, that the Apple one goes out for its work. Anyway, yeah, it would make sense these days that it runs on chip. Okay, continue. Yeah. So we've got sort of demos. So I've got a sample file here. Choose a model. So those are the four models that are available. And it's a trade-off. The larger models are slower, but they're more accurate. Yeah. Though I would say that with the large model, if you're tight on RAM, that you can find there's a point that the large model actually is worse than the medium model because it's when when things run tight on ram you start to get uh because ai models are all very ram intensive so then just try it right right right so the first time is a bit slower because it needs to load the model but that's 100 local and i know it's a tiny little
file but yeah so this is like to pause about a couple things um i've run a lot of local models I have 18 gigs of RAM on my MacBook M3 and 16 on my MacBook M1, which is my open clock bot. You can't run a very powerful AI model. I mean, you can't run anything that's even close to like Cloud 4.6, not even, I mean. However, for transcription and also for like encoding and for certain embedding for certain types of AI functions, you can run a perfectly solid one. And that's clearly what you're doing here. So the biggest model, how much RAM does the biggest model require? and processor that's a good question and had i been prepared i would have told you exactly well i'm going to play with it and we'll talk about it on our next one when you talk about the other tool but i mean it runs comfortably you can run it on your computer yeah yeah so this is um my this is my main computer it's a yeah 16 gigabyte m1 yeah there you go if it runs in that then there's no problem it runs in almost everything that's essentially what the newest There's a brand new little Mac, right?
The new really sweet $500 whatever. It's eight gigs of RAM. And I don't think you can order that anymore. I would hazard a guess that the large model would be too big for it. The medium model might be a stretch. I think it's really interesting that in the world we're in, that actually Apple is coming out with machines with less RAM. maybe that's because of the resource price of ram right now uh anyway that's what i've seen the performance of those should be quite good if you're on something like one of the new say like an m4 m5 oh yeah hell yeah it's yeah gonna it's gonna be just screaming fast well that new little mac is actually an a series chip not even an m series so it's like a chip borrowed from an iPad. Anyway, continue. Yeah. So, it's quite -- is that the transcriber folder? Basically, this sort of looks at the documents, the process.
Drag and drop, point files. I say returns JSON with complete transcript in the JSON, which is exactly what I would be used to. Sorry. If I just clear all of this. If I come back here, where I'll just-- so what I've done is I've specified the folder. As you can see, it's Documents Transcribe Batch Folder. Yep. If I grab a few of these things, drag them across. Oh, hang on. OK, I'm curious-- I want to keep the copies in my samples folder as well. Duplicate, sure. Yeah. And if I just run that, what that's done is it returns immediately a job ID, tells you it's queued, and that we've got four files. Nice. We've also got sort of a status endpoint that can tell you, you know, sort of how the process of this is going.
What I've got it doing here is I've got it polling every 10 seconds to check and see if the status is completed. It's not setting that value here, but it's going to take a couple of minutes because what we've got there is probably about two hours of audio recordings in those four files. I should have got some smaller samples. No, no, no, it's great. So this gives me a minute to ask you some other questions. So this is transcribing like audio of people talking. Is it just English? No. It's multilingual. Okay, cool. And then what does it do if there's more than one voice? It will transcribe it. The version 1 does not separate the speakers. Version 1.1, the one that's coming real soon now, does separate the speakers.
Or rather, I should say it gives you the option to separate the speakers. Okay. Because in some cases, you might not want that. I mean, in a lot of cases, you would. and i'm also super curious like how many how many speakers that'd be like uh generally these things are pretty good up till about say six or seven then they start getting a bit confused got it so not a not a city hall kind of a meeting with like you know people testifying every three minutes or something like that but i mean yeah normal conversations podcasts actually this is totally my interest rate i can drag my whole podcast history into it uh yeah i should have more in my folder than i do because i've been a podcaster ian for 20 years believe it or not 20 years wow i know that's that's quite impressive it's long the time flew okay so okay so what that's done is that's come back and it's told us so you can see the durations here in seconds sorry the the demo file is actually what i'm not yet finished so that's why uh that's why some of the you so you can see that yeah we've got you know 1900 seconds that's about half an hour that one's about two-thirds of an hour another one you know 25 minutes or whatever 11 seconds for the jfk one
yeah so yeah so that's kind of gone and and if i just uh and i'm sure you could you could change Is that to be like hour, minute, seconds or something? Absolutely. Simple calculation. And then what we've got are the... Transcripts, yes. The transcript results. So the demo file will be finished. So that's, yeah. On the process of time. Yeah, all the stuff that I would expect to see, right? When it was created, what the job ID is, where it is, what the entire transcription text is. The other thing, actually, that that sample doesn't show is that what I've done is I've given the ability that if you're doing a folder, if you're either batch transcribing a folder or you're watching a folder, you can also assign metadata.
So what that means is you could say, for example, you could assign a source or a job number, stuff like that. Right. Okay. Which, again, then all comes back as JSON and into FileMaker. So it's trying to return everything in JSON format. And FileMaker just works beautifully with JSON, so it makes it really straightforward. So what if the audio is garbled, or the audio is really bad, does it-- I'm curious what it does. This is really more of the model than your use of it. I just kind of wonder what happens in that situation. Yeah. It varies, to be honest. And in a couple of cases, when I've had the door open, when I've been testing, I've seen it sort of pop in something saying bird song or music.
Or it grabs the audio from the TV in the other room. Yeah. Actually, when I was testing it, it was quite funny because I had this one time where I was testing the recording and it just was not working. I was having all kinds of issues. And then when I eventually traced it down, it was around the level of the volume. And I discovered that what I'd done was I'd left my headphones on, and they were the input source. And I got them stuck inside the little case. And the case was the other side of the desk from me. Got it, got it, got it. And so it was all quite muffled. It's only going to do so good of a job. Sometimes these technical problems you try to solve have very simple non-technical solutions.
yeah for sure yeah um cool so yeah i'm kind of thinking i don't know part of me is thinking like how i'm going to use it because i am um and like how also this would be deployed in production so that since this is a mac app obviously the database that it's writing to can be hosted so the the app would run locally it runs on a mac very few file maker servers are max and shouldn't be, in my opinion, unless like if it's a local box that's doing other things, frankly, if it can actually do AI, if you've got like a Mac studio with a ton of RAM. So you'd have like what we used to call a robot machine, like a Mac sitting. That would certainly work well for things like the batch processing.
Or a watch. Alternatively, what you could do is with the Linux version. So the Linux version doesn't have a UI like this. But that would run on the server. It essentially is just... That would run on the server? Yeah, it could do. You could run it on a server or on anything. I've actually tested it on Raspberry Pis, and it's admittedly slow. But if you think about it, if you're watching a folder and you're doing stuff that goes into a folder, then often time isn't quite so much of the essence. Right. If you're doing it locally, you want it like, you know, I want it now. Whereas if it's something that runs in the background and people just drop files onto a server, it can do it in the background and leisurely.
The other thing that I'm adding in version 1.1 is webhooks. So when it's done the transcription, it basically calls back to webhooks. yeah yeah yeah yeah oh that's cool so like uh like the flow i'm thinking then is um you've got a google drive or dropbox folder the app is watching it um and then a file gets dropped into the folder it gets transcribed a web hook calls a thing that posts something in a slack channel or something saying it's done you know kind of agentic exactly yeah or back into file maker so you could bring the you can pull the data back into filemaker directly with the web hook if you wanted to yeah or like make a record anyway yeah yeah that's got that's good this is very very purpose like very single purpose focused utility uh yeah so the the the second one is it works the same way in terms of the process that the folder processing well before we get to that where can people get this and when um actually i've got i I've got a page on my website that I thought I would have given you in advance.
No worries. What's the site? Okay. It's transformingdigital.ai/transcriber. Transformingdigital.ai/transcriber. Without the E, right? Without the "e"? You should maybe make a link with the "e" so people can get it. Okay. There it is. Species, texture, mile maker, 100% local. Notify and see how it works. So shortly there will be a download thing, yeah? Absolutely, yes. Okay. Yep. So the second product is essentially very similar in terms of the structure of how it gets data out of folders. but what it does is extracts document data in a structured schema oh so you basically have what you would do is you'd hand it a pdf document or a word document or whatever and you give it a schema of the data that you want out and it gives you back the data that conforms to that schema 100 so this one also is super interesting to me because i have another project that has immediate
I have a lot of production things that are AI, and this is one that really hits me because there's tools like FileMaker has a tool that can get a PDF and photos and things like that, but there's not an easy way to do Word or Excel at all. Well, I would also say that what FileMaker gives you in the PDF is it gives you the text. Just the text, right. And the text is not what you want because the text out of a PDF could be in any format, anywhere this will give you the data back in a structured format yeah so the files endpoint will um will actually give you an easy way to upload a pdf file you know easyish and then you can do a generative ai call that says give me all the information kind of like what you're doing maybe that's what you're doing in the background um but it does not work with with excel or word there's extra lifting you have to do there's filemaker plugins for that but i don't think they're going to be structured like what like what you're going to have here so is this also entirely local yes yep entirely local cool and um yeah it's actually really quite cool because it can get you quite complex structured data out really effectively how do you provide the
structure what's that look like what's that look like yeah it's a json document so like if you Like if you have a bunch of invoices in PDF formats or products list up, you know, price list, update spreadsheets in Excel or things like that. What other types of documents are supported? I'm not guessing. Well, essentially PDF, any kind, any, any image, Word, Excel, PowerPoint. Oh, and photos. Yeah. Yeah. And photos. What else? Yeah. And it's yeah. So, so in terms of the structure, you kind of tell it what data you want out and in which format and it will give it to you like that. Nice. Okay. So in the case of an invoice, it will extract the data structured as you want it. So, you know, you can specify, do you want just the header?
You know, do you want exactly which bits of information you want? So it knows exactly what to get and what to ignore. That's right, yeah. Beautiful, yeah. These are things I teach in my free classes every week. So sorry to keep interrupting. Yeah, so, and now I'm going to be showing this tool and then talking about the kind of stuff that goes on under the hood. I love really, like, seeing and teaching what's happening under the hood for things like this. yeah so yeah um now there's one other thing we touched on really really briefly and that was that when i was writing the demo file i had a bunch of file maker scripts to write and that got me off down a rabbit hole which delayed this a little bit because as these things do um what they do is i thought well actually i wonder so i've got And I've got a load of scripts.
Yeah. Health check, law firm info, transcript file, list transcription, the sketch transcript. Yeah, okay. Tools. So one for each. However, so what I did was I started off with these scripts. And as you do when you're writing FileMaker scripts, they were a little bit, they weren't all consistent. So I thought, actually, can I just write this with AI to simplify this process? So if I copy the script text. Actual script text, yeah, from within Script Workspace. Well, sorry, within Script Workspace. Yeah, yeah. And if I go over to Claude. So you're in iTerm? Yes. Oh, ClaudeCode. But you're in not just regular ClaudeCode. Sorry, okay, I live in ClaudeCode in Terminal.
That is where. Got it, got it, got it. That is where I spend 75% of my data. So walk me back on this one because I don't use this. Okay. I'm sorry. Well, how do you use Cloud Code in Terminal? So I can start doing that. So I've got a Terminal window here. And I'll just say pwd for present working directory. It shows me I'm in this FileMaker Claude folder. And if I just, I'll just open that in Terminal. You can see I've got, it's just a regular folder. So you type open dot? In Terminal, open dot will open the current directory that you're working in, in the Finder. Not on mine. Really? Yeah. open space dot oh no i didn't watch exactly what you're doing did i yeah so that does it opens up the the folder and so you have a you have like a cloud folder so you do cd and get to the folder that you want um yeah so actually what you can do is you can go to any folder on your computer
and then in terminal type cloud if you've got cloud code installed type "Claude" and it opens it. Command not found. So you've got to install Claude terminal version, not just Claude code. Is that done from within Claude? Claude, you do that -- Probably just -- You probably did it a long time ago and forgot, huh? I think this kind of thing is a really good aside for the podcast actually, yes and I'm sure it's going to want a bunch of other things brew, I knew this was going to be it so brew install node oh, hang on that's not the specific thing No, but that gets you Node. Brew install cloud code, maybe? Hang on. I'm going to just do it while we're talking.
I already had Node installed, but I'm going to run it to see if there's an update. Node is huge if you have OpenClaw. It uses it constantly. It's shooting stuff in the background. Yeah. And then, yeah, I mean, even though it's a little bit of a pause, it's not too long, everyone. There we are. Okay. So -- oh, they've actually got a native install that they're now recommending. Oh, actually, there is a brew. Brew install. Brew install -- Dash, dash. That's just cask. Dash code. So it'll probably do the whole thing. Yeah. Brew install. Here, leave that on the screen for a second so I can go do it. Well, I have to finish for, okay. So anyway, so you live, while I'm waiting to do that, so you live in Cloud Code and in Terminal.
What's the benefits, by the way, of that versus just Cloud Code desktop app? Cloud Code desktop app can watch a folder, so you can have a folder for each project. Yeah, so essentially you can. And then it can read and write it. I've just found this to be quite, I find this to be sort of more usable, to be honest. I do have the desktop app installed. I tend to use that for questions that are kind of not related to what I'm looking at. So I've got a number of different project folders. and this enables me to have my multiple project folders all open simultaneously with different instances running. So I've got... And you're using iTerm, so you're not using Apple Terminal.
Yeah, it was a little bit more configurable. Oh, do you still need that Safari window? No, I actually have it installing right now. So just to repeat it a little bit, is brew dash or brew space install space dash dash cask as in of amontillado space claude dash code c-l-a-u-d-e not c-l-a-w-e-d yes and we'll put all these in the show links too okay yeah or the recommended approach is like a curl yeah which is interesting so that that's interesting so native install recommended yeah curl f i'm not going to read that one right uh no but it goes to it points to claude.ai and so you know and you probably can get a link right from there and get it yeah i would expect so windows powershell windows commander mac or linux yes okay cool and then then the login stuff so and then it logs into your to your um to your Anthropic account?
That's correct, yes. Now, that was something that initially I found-- Back to how you use this to write FileMaker scripting now. So you have some FileMaker script steps in memory. Yeah, so I've got the FileMaker script steps in memory, and so I basically-- and it took a while to get to this point, but-- so here is a-- Here's a FileMaker insert from URL script, so you describe what it is that you have. While you're typing, I'm pretty sure that when you copy script steps, they're not pasteable anywhere. Oh, yeah, you have the wrong thing in memory. So they're not directly pasteable. other likes uh it's a it's a i can't remember exactly the correct wording but it's oh uh this could be an mb this could be an mbs plugin that's what i thought yeah mbs plugin does have that feature when you copy script step code to the clipboard it puts it into a format of xml that can then be directly pasted in as text with all of the information so like if there's a long
calculation it puts the whole thing in as opposed to just the little part of it that could see which which is what otherwise it used to do. OK. So I've pasted it. And you can see Cloud Code is telling me pasted text. Yeah, without giving me every single line. Yeah, OK. Yeah, so let's run that. And that basically expands that. You can see there's my script. Yep. It goes off, and it's pondering. And various other things. Yeah. I love how it changes the words all the time. yeah this will take a moment now the thing is i've actually built i built a cloud skill to help me with the insert from url script step because i had a load of them so you can see i've got a skill fm insert from url it's loaded the skill automatically yeah and i've created i've kind of refactored it a bit so it's now got two skills there's a a FileMaker base skill so it knows about FileMaker separate from the...
Because I had it all latched into one. Not to get too far off, but how do you define skills in Cloud? I was a super open AI guy for a long time. And so I kind of know how to do it there. But how do you configure skills in Cloud easily? They helpfully give you a skill to write skills. Oh, there you go. So if you want to create a new skill... In Cloud Code, you just call it... So, anyway, it's -- Okay. So, it's giving you information about the code it just read. Here's the assessment of my script. Mm-hmm. What's it doing? What are the issues? What's working? What's working? Issues found. It's basically no error handling, no error checks. Are you good with she? So in my skill, I've asked it to add the show error just because to get more information.
If there's an error, the dash S option gives much more information to debug what's gone wrong. Do you call Claude a she? No, I don't. I thought I just heard you say that. Anyway. Oh, the missing. I asked Claude what pronouns I wanted it to use. I recommend people do that. It was pretty great. Okay. Nice. And it's kind of ranked by severity. And so the URL not decomposed. Okay, cool. And then so you're using this to sort of, yeah, you're using this to troubleshoot and check. Then what? And then it basically says it can copy it to my clipboard. The fixed version or the version that does new things you want to do. Yeah. I'll just go into a little more detail on how this does that.
Okay. Yeah. While it's thinking. Actually, once it's done, it won't take long. So it's remaking XML that specifically works with FileMaker's somewhat brittle XML script step format? That's right. He said carefully? Yes. And there's a number of developers who are working on this part of it. Absolutely, yeah. Yeah, there's several. I think Aggie's got a product. She does. It's great. Yeah, but it's way more than this. It does a lot more than this. Matt Patrowski has some. It's on GitHub. Yeah, yeah. Those are kind of the general purpose. What I was looking at here was, I mean, this is a very... Okay, so yeah, because I've done different versions. Yeah, and Christian, of course, is working on this directly in the MBS plugin, because that was where I first saw it.
Just like we saw a minute ago where you can copy it, He also has the ability to paste it. But it didn't work-- at least in the first versions of it, it was maybe 20% or 30% of the time that the paste worked because of that brittle XML format. OK. Yeah. So what I've done is essentially it's created that. Mm-hmm. Batch Python, change, recompose, added layer, added this, added that. Oh, yeah. So you pasted it. It gives you all kinds of comments. wow beautifully commented uh and then all the code pasted in without error referencing all the correct fields and variable names and the correctly escaped and formatted uh curl commands wow yes this is the time saver this is kind of what well this exact function except without the back and forth copy and paste is what is needed in FileMaker.
Yeah. Without question. Now, the other thing I can do here, and this will take a bit longer is, so in my folder, this is it. You'll see right here, Swagger doc. Yes, Swagger doc. Yeah, yeah, yeah. That is the Swagger documentation for my API endpoint. Oh, okay. Oh, so you just give it a command, look at the Swagger doc. Oh, and then it knows how to configure it for all of the specific JSON in that file. Was it an MD file? Or JSON file? They're JSON. In this case, the Swagger doc is JSON. Yeah, okay. Now, this will take a bit longer. So what's in the Swagger doc? This is... i'll just uh we're super looking into the looking under the covers here i love this a lot so the swagger dog basically swagger doc describes the uh the end points in json format so for each endpoint because you've got the host the base path what paths are supported what method to get that's a get this is a post so the batch process
directory is a post, description posts all files in a directory. Yeah, so it's a function library of the API call, basically. Yeah, exactly. What parameters, what responses. So that's really, really handy. But that -- anyway, this will take several minutes to do this. However, if it takes four or five minutes, It's going to write a FileMaker script for every endpoint in that API in FileMaker format. So a huge amount of code you then get to copy and paste. What's your success rate on this? This is actually a whole other great tool. Okay. The success rate on this one is, at this point, I'm going to say it's pretty darn close to 100%. The only thing that this one's going to do is that it's not going to know which fields to set in setting the results, the set fields. So, it's going to leave those blank, but it's going to stick a developer note above them saying, "Developer, you need to change this."
Of course, MCP flashed in my head there for a second, where you could have an MCP function go discover all of the structure in your FileMaker database and get, or even an API call, even a FileMaker data API can go surf the database and see all the layouts and all the other stuff and then know all the field names and layout names. Anyway. Yeah. Anyway, that's working way now. What are the things that I or other people would need to be able to start using this function of being able to copy and paste script code so it looks like this is going to work either with the regular cloud code or the cloud code terminal version yeah but what's the magic and then we already know the one bit of magic which is copying what using the mbs plugin which everyone should own and buy um even though i think this copy function comes with the free version uh yeah and then so then once it's pasted in obviously that's going to be working but what's the what are the other missing elements i saw you had your special tool to then return from your prompt code you can paste back into filemaker script workspace yeah what would you say
so let's try it cold and see what i get but like in other words did you have to teach it about that yeah the brittle xml format somehow somewhere yeah i i basically gave it i basically found i did a load of research and found the that format yeah which i would guess would be public shareable yeah so i i essentially i did a number i gave it several examples where i copied and pasted right um and then let it figure out what was yeah yeah i mean you could give it examples of you know hundreds of scripts with all of the from whatever app whatever applications then it would then know what uh script steps exist what parameters they require what the format of the parameters are the weird ins and outs So some of them that require quotes and not.
For example, when you use a JSON function, you can use JSON string not in quotes. But if you use something like get container attribute, it has to be in quotes. So there's little details like that. Yeah. Oh, hang on. a really long time ago in my early filmmaker career i took a bunch of eight ux classes when apple had that you know like like right when os 10 came out god when was this 80s uh yeah the company i was working for needed to get a bunch of system engineers um uh certified in that so they flew us all over the country for at apple places to get these classes And that was where I learned all of my Unix commands. And it's been a long time going in. So I don't use it.
You're asking, where does it know how to do this? So I've opened up the .cloud folder, where it's got the skills. And so I've created a skill, fm_insert_from_url. And if I just double click on that-- Fm_insert_from_url.md, markdown. So it's basically all described in a markdown document. So, basically, description, and insert from URL, and error handling. So it basically -- Is this on git or something you'd share? It's not on git yet, but I'd be happy to share it. Yeah. I'll put it on -- I'll put this one on git and share it. So essentially, what we're saying is we need to have a couple of layers of error handling. We need to check for connection errors. Then we've got API errors.
So in front of me, this can all be adapted to the way somebody wants this stuff to work because people may not like the way that I've got the comments done. Everyone's got their own. Well, I mean, this file is also something you can just open in Cloud Code and say, do, you know, do this with that. And every time a new version of Opus comes out, you know, give it another look again. But this, by the way, that leads me to another question I had, because, you know, being a person with ADHD, my brain is quite nonlinear. And maybe most brains are, I don't know. So what happens when you have your compiled download Mac app and a new local model comes out that does transcription better?
I would basically need to recompile it as things stand because the list, the dropdown list, is built in. So what it does is it ships with the base model already because the base model is not very big. It's a, or is it a tiny model? Yeah, so it doesn't work. It ships with a model. The first time you run it, what it does is it says, they've asked for, say, the large model. I don't have the large model. It goes off and it downloads the large model. Ah, I see. Oh, so like at whatever time that could be a different model or is it just like a kind of hard-coded specific one that we know works? Those are currently hard-coded. Well, it makes it much more reliable if you do it that way.
Because if it downloads one that's broken, then your app is broken. But since the app doesn't store anything, if a new version comes out, you just go update the new version just like anything else, you know? Like the Zoom app. I updated Zoom today to version 7, hoping that the new features would work. And here they are. Oh, actually, this is not told me. Okay, we're done. Great. So now this just updated a bunch of scripts, yeah? It didn't update them. It actually wrote them. Wrote them. Oh, man. So if I create a new folder. Oh, one interesting thing here is that if you start with an empty FileMaker file, you need to create at least one script before you can paste anything in.
Yeah, or like a folder. Ian, this could have been the whole podcast. I mean, the tool is amazing, but getting at how you made it and how you're updating it, this method is a game changer to a lot of people. And there's one thing you referred to. Yeah, I see a set field where it says, hey, FileMaker developer, fix this because I don't know what the context is to the field. Yeah, exactly. Now, the thing is, if somebody had a standard thing, they could actually put that in the skill. Or to be honest, you could, I suspect, I've not tried it, but you could probably say when you ask it, if you gave it the reference. I'm not sure exactly how that reference would look if you typed it in text.
But to be honest, that's the only thing. And I think that's just amazing that it can-- Yeah, hugely. I'm going to take another quick aside. I'm looking at your script code, and I see something that says, output to field, developer note. Set the target field for this thing with an icon. So MBS has another function that people might not know about, and that is a search function within a script workspace that searches for every single script in the script workspace by script step or variable name. It's right here. Right, right in the top right there. So you can actually type, like if you, you know, you have, if you're replacing, for example, you know, the send mail script with a better way to do that, you can just do a search on that script name and it will list every single script that calls that function and then you can go fix them one by one it is a game changer um like a lot of the things we're seeing here today um anyway kudos to christian yeah absolutely that's um that's really pretty awesome that's yeah which i'm sure you're using and that way because like if you if you save this script right now and you close it when you come back in it's not highlighted anymore it only gets highlighted like that first time you open and look at it um when you when a script is made
um that there's something that you need to to address yeah so what's really interesting is that i mean you keep you keep hearing that everything changed around about november last year when yeah yeah when when anthropic released claude 4.5 and now 4.6 That feels like that was like a big hump that if people looked at Cloud, and I'm saying Cloud because that's what I use, but OpenAI is the same, Gemini is the same. They're all the same. Around the same time, there was a big step up in capabilities. And the other thing is that prior to that, we'd always say, I wish these things knew about FileMaker. So they're not perfect on FileMaker. They do make some mistakes.
But if you're on one of the paid plans, by the way, the paid plans give you better results than the non-paid plans. Sure. As you might expect. Bigger, faster, better models start at $20 a month. Do you have the $200 a month? Yeah, so I'm on the, I'm on the, I'm on the, I don't actually. i'm getting by on the hundred dollar a month one oh okay it just it's i'm kind of close to limits but i'm not um i read a blog post from someone who's doing um they've got a structure they refer to as a software factory that has like a lot of parallel agents right yeah each one of them with their own 200 a month account yeah yeah and he said that apparently he triggered anthropics fraud detection thing because he purchased something like uh 50 or 60 200 a month accounts and anthropic thought this was a suspicious behavior sure and he basically no no no no i want to buy more yeah um so yeah i mean as expensive as that is you know so like each agent can kind of replace, I don't know, like a $30,000 a year or whatever employee.
Not quite, but I mean to some degree for $200 a month. At least that's kind of where we're going. But the other thing is that this can actually be used to actually just check the script. So if your company has got, say, a standard for a particular type of FileMaker script, so insert from URL I chose because it's a little bit annoying that there's a lot of fiddly bits to it. There's a lot of things I don't like about it, but it is our way right now. Absolutely. It shows the dialog box, which I hate. And the format for exactly how you have to do it when you put it in is weird. It's definitely a command that Claris had no idea would become the focal point. Now, can you use run response, perform response from model, get response from model?
For this stuff? I wonder. Well, in the new method, there's a new script step, right? You can configure the AI account and you just name one. And then you can use getResponseForModel where you can send your prompt there, which works for local as well as all these other ones. Yeah, the problem is that your prompt has got like a lot of stuff. your prompt has got all this knowledge encapsulated in it not only that but it also knows which tools because in there supposedly those are all configured i i maybe it's lazy but i've just relied on insert from url because those things are so much easier to see and control but the get response from model has all of those checkbox for like use these tools use this this here's your system prompt here's your user prompt blah blah blah yeah oh this by the way is the stuff that it created when it was working on writing this ah okay created the folders in your in your desktop well this is in the skills in the cloud skills but what it did was it it when i was writing the skill because it knew i was scripting it needed some kind of evaluation so evals so it wrote it asked me for some evaluation so it it asked for examples of a get script
a post script and a delete script um so it did like all the evaluations when it was building it and then it sort of built up these things like these benchmarks while this is while it was creating the skill you can write the skill you can write the skill manually but to be honest if you work with their tools to write the skill it will be quite and then you can see the tests that it's done so what it did was it compared it compared like running it with the skill without the skill nice so here is the prompt we're going to write a FileMaker script from scratch close delete that end point all of this blend of using AI to be a better FileMaker developer to learn stuff and just to make the things you can do vastly more powerful the speed with which you can do them vastly faster as well as deliver results that directly benefit our customers is so transformative.
I'm really focusing on the latter and teaching all my classes on the latter. Yeah. I think there's a lot of places where it helps us. Okay. Even if it doesn't allow us to totally write everything in it. Okay. My opinion shifted actually quite a lot in the last probably six weeks i've i'd been quite concerned about clarison filemaker and what's going to happen i mean i still am but my my experience with this and i've also used the i've also got another example where i used the fm upgrade tool um to and what it'll do is it'll actually create a schema right um oh do we have time how much time do we have uh some time been recording an hour sometimes these go an hour and a half but you know okay we don't really have a i'll just i'll just um i'll just go over to my um this is my mac studio and i'll just show you an example here so i'm going to create a new file maker file
Let's call it Matt. It's absolute minimum FileMaker file, just as they created. The only change I'm going to make is to give it a password. The reason I'm going to -- yeah. Just make it admin, admin. The reason is that Claude gets upset when you run commands with blank passwords and thinks you're trying to bypass security. So I'll just close that. I'm in a different instance of Claude, different computer. So this is, yeah, this is like a 32 gig. It's a Mac Studio M2. Yeah, not the crazy 512 gig version you can get. I kind of wish it was. Yeah, I mean, you could run really big local models on that one. Yeah. And apparently you can gang like six of them together with a complex array of connecting the cables from one to another in a specific order.
Yeah, the software that does that will actually work across a network as well. I tried it some time ago, and it was actually the software to make it all work was really fiddly, and I gave up after half a day of... I would bet. I'll just give it this, just so it knows what we need to do. Okay, I need you to pick an industry. Oh, actually, no, I know. So, you give it a prompt, like, made this new file. Here's the credentials. Yeah. I want a simple -- Yeah, this is your prompt. Do this thing. Yeah. Here's the industry that I'm in. Yeah. Yeah. This may take a minute, but we'll probably come back. And then what we're going to do -- yeah, this is --
This whole thing we're getting used to compared to six months ago or whatever is before, when we used ChatGPT, we would never wait. It would always give us response in some seconds, right? was kind of nothing we could ask that we wait for for five minutes or a day and now we have this feature for it to do much deeper thinking and solve much bigger problems and maintain a folder of local directories and point it at git or point it at a folder or whatever okay so this is created then well it's actually not done it's not created it yet but what it's done is it's come back and said here's roughly what we're looking at and then says okay want me to turn this into a file FileMaker patch schema for?
Yes, please. It's FileMaker-- what kind of schema? A patch schema. OK. So this is then using your upgrade tool, yeah? No, this is not my upgrade tool. This is the upgrade tool that FileMaker has provided us with, the one that has been so complex that nobody can really make it work very well. Actually, Simplice Alliant did a really, they've got like a reference tool that like has explored all the different patch formats. It seems to be a tool that's kind of like 90% complete. But does it take XML, if I'm remembering correctly, and then converts it into not only the script workspace code that we saw earlier, but also everything, fields, tables, layouts, scripts themselves.
Yeah. So the idea-- the original idea was that what you could do is you could write your FileMaker file. You would sort of-- the process, I think, is you generate GUIDs on a FileMaker file. You then make your changes. You save as XML. You create a diff of the two in a patch format. And then you use this FM upgrade tool. Yeah. Interesting. And it's going to ask me to do a couple. This is really, really all very, but you can see it's running the FM upgrade tool, sub command, generate GUIDs. Yeah, which is the terminal command, of course. Yeah. Yeah. And it kind of, well, I can't say I wrote it, but it wrote an FM patch generator. It doesn't do all formats. It's very much.
So it's actually done, but it's going to now look at the log file to make sure that it worked. Wow. Created eight tables, 68 fields, 10 relationships, eight layouts, and the partridge in a pear tree. Yeah. Wrong time of year for that. And then open up that map file and drumroll please. Boom, wow, a ton of layouts, a ton of fields. Is there a graph, relationship graph? That is the cool thing, because previously, like with OData, you could do all that with OData. Well, I mean, you couldn't create the layouts, but. Wow, they're stacked, but it actually created the relationship graph with the relationship with the correct keys. Whoops. That's kind of magic right there.
Okay, it's a mess, but... Yeah, yeah, but it doesn't matter. You can just go in and move stuff around. Oh, and it's spider graph too. Isn't that cool? Yeah, well, yeah. I mean, because, well, yes. So anyway, but isn't that amazing? That's totally amazing. Essentially, you've created a complete FileMaker database from a prompt. Yeah. So it's very... Yeah, probably not much. I don't see that the layouts don't have UI yet. No, they don't. And that's another... I mean, okay, look, this is basically from, you know, days playing around. And there's a lot more that would need to be done here. Yeah. So my next podcast I'm going to record is actually with Stathis, who has a method similar to this.
Yeah, he showed it at a conference that basically creates the whole UI entirely in a web viewer. Yes. Actually, I'm not sure if you've seen his demo of that, but it's amazing. so marry that to this and basically you have you know prompt a create the file maker database with all these fields relationships layouts blah blah blah prompt b go create a ui um with with all of that information knowing what this app is supposed to do and then it creates that ui in javascript code that runs on a web viewer and poof you have an app i mean yeah this is this is stuff that you have you know you've been able to do all this you can do all this without being involved in all but But I still really think that there's something great about having FileMaker, either an existing app, or just the data in FileMaker tables, and the FileMaker server involved, that gives you so many other capabilities, not just to connect to legacy apps, but there's a lot of other features about that that I think are really, really beneficial.
Anyway. Yeah. It's really not the least of it. the first test I did was that I just, I basically just sketched out on paper what the ERD would look like. And did you scan that and bring it in? That could be part of your config, couldn't it? Yeah. So hang on. Did they have the invoices contacts? Let's say I added the schema with the structure and a photograph. Please look. Actually, what I can do is I think I can just drag it. Can I? Oh, nice. Drag and drop into the terminal. Yeah. Cloud code does that. Okay. So, from the diagram, I can see that. so this the the other part the the part that does the the script that you paste in is is much more focused on a very single use this is kind of still at this point this is really
like early proof of concept stuff but the fact that it will work nicely with the um the upgrade tool I think is very, very impressive. Yeah. Amazing, actually. And that clearly is the magic required for it to actually create a FileMaker database because it can't otherwise do that. They realize that, yeah, there are some issues. But it also knows, because it knows the structure, it knows that it can use the FileMaker developer tool which i didn't tell it about actually i did not tell it about the developer tool but it worked out that hey not only have you got the upgrade tool you've got the developer tool here and i can use the developer tool to save as xml and because i can save as xml i can look for the problems in what i just added like if i but the other the other thing it was doing was it was it was writing changes and then it was reading the log file but it wasn't trusting the log file 100 it basically would then go save as xml and then check the xml to see is everything that i expected to see there is it there so that's kind of cool uh so that's done
let's just save that okay so next round contacts facilities no sample data but I'm sure that can be done so did it just it just made more relationships based on yes based on that photo that you put in yeah damn This is a really nice-looking graph, Ian. Isn't it? Add another couple thousand tables, and we'll be looking like one. We'll be right like all the applications we've seen for most of our career. But, I mean, if you wanted to, I guess it would be some work, but there's tools that we know of in regular FileMaker land to take a spider and break it out into Anchor Buoy. uh to be honest the way this is working i bet i bet that you could do that i bet you could teach it what you want how you wanted it linked up and joined up the problem is that everything kind of showing up stacked on top of
so you might end up in a situation where it creates all the relationship but then you need to spend an hour rearranging yeah well i wonder if you just did this in step by step you could have it not create the whole spider but you can say you know like the thing that you just did where you said scan the document you could say don't connect this to the existing tog you know make a new yeah i would expect so yeah yeah and then you get get your feature except it might it may or may not refer to existing tables so that would be well these are more things to play with right so the whole of ai the whole of what we're doing is a balance of using the tools and spending time to learn as humans the best way to make them work because there's there's still a lot of intention that you need and a lot of skill required in using all of these things for the five percent of the stuff they don't yet do and to try to figure out that part that's where i spend a lot of time yeah but realistically if you think about it sitting down in file maker and and typing out a hundred fields or whatever the number is right is that's
not where the skill is where the skill is knowing what is the business problem we're trying to solve the business problem to the technology how do we what fields do we need um that's where the that's where the skill is I kind of here clearly took a shortcut saying just research reference data schemas for oil or whatever of course which is the right way yeah because that's a quite good job of coming up with the right ERD for something like that a pretty good starting point I suspect yeah yeah yeah hugely yeah I'm doing for I'm doing a Dropbox API integration which I haven't done and I spent there's a couple of hours of round trip of getting like the oas stuff set up getting this pasted in getting a hello world test getting this you know there's a lot of round trip back and forth of it's not working i get this error oh well that's because of x well actually no that's not what was wrong so there's a lot of that you know ai confidence of like this will solve it and then it doesn't solve it um that i'm sure we all experience i hope we all experience uh and it's it's not just me yeah yeah no that's that's very much the case and yeah and i can see these are us
learning ai's coding because it's it's highly confident goes off and and there've been several cases where essentially and it's quite funny because with this stuff with the the fm upgrade tool at one point it literally became it had convinced itself this thing does not work it just kept saying once it had got it in its mind that this this aspect does not work um i ended up having to clear the context you know you know close the session start a new session try again and it very quickly came to the same conclusion and and i got it i basically had to prove it i i said look i know it works because i've i have people are doing it yeah and and i said there's the FileMaker upgrade tool was designed for this very purpose.
And then, yeah, sort of then I fed it every bit of documentation I could find on the upgrade tool. From Clarus, every blog post that I could find, you know, from a bunch of really smart people who kind of wrestled with this. Sure. Yeah, and I love that our community has been really great at sharing everything here. Yeah. Yeah, I think it was Salient had some really good blog posts on the tool. And what was it that they – I can't remember the name of the – there's a FileMaker file they've got in GitHub that essentially is a research tool into the FileMaker patch format. And that – I haven't played with it much yet. going to patch lab that's what it's called okay i have done more learning and trying things not only just with filemaker but with ai in general but i wouldn't even have dared uh you know a couple years ago and it's been so satisfying to be able to do all these things and then i'm working with other people that i know that had different previous skills coming into it and feeling
like an idiot based on the things they're doing but you know you can't have to just put that aside and do what i'm doing what i know to do and focus on my own joy and my own uh you know um things that i'm delivering and i'm also i'm doing something that really not a lot or nobody else is doing which is putting these free classes together every thursday and sharing everything that i know you were in a class i taught last week and i discovered there was something that i've been doing that's completely wrong and i figured out another way to do it which is another little adhd aside which is trying to rely on a generative ai call to do math even simple math where you can say here's an account in a structured json with 50 invoices go add them up by product type generative ai is a token prediction engine it does not have the smarts of a calculator from 1966 it doesn't and you can try to do a tool call but it's just not going to be 100 reliable and so what it does do what i'm going to be trying to do is have it write um like a json block or code or some other type of thing
Have it write something that is reliable and have the AI convert what I want and send the data along unchanged, which it can do, to this tool, which then can slice and dice it and then display it in the HTML format. That way it looks the same every time. It's a different flow. It doesn't quite feel as satisfying, but it delivers a better result. Absolutely. And a lot of these things, sorry, these things like with AI doing a simple arithmetic, and I remember people holding that up as an example. It was really bad before. But then if you think about it, if you asked a really smart person to add, you know, two 16-digit numbers together or multiply them without a piece of paper, they're going to get it wrong.
Yeah, but again, a calculator from forever ago. Yeah, to do this. Yeah, but also I found that if you use the files endpoint and you upload a huge document that has that same exact JSON structure or just the raw data and multiple tables, and that could be massive, you know, close to the context window of a million or whatever. and then you do a call on that that actually runs a whole different process which only uses a tool or it writes a tool and uses the tool in one fell swoop and then gives you a response which is exactly what you want yeah it's a different path inside the ai but yes since you just mentioned context i'll just show you something in cloud code here that actually, so.
Yeah, then we should wrap because at this point only Alexia is still listening. Oh, okay. There's a command called context. And what that does is that shows you the usage and it shows how much the, you know, we've got, so we've got a million tokens, the system prompt is six and a half, system tools, memory files, skills, and then you can see the messages down here that's what we've passed back and forth you know i don't really use these slash commands a lot but damn that's great i just tried it in the ui version too it's great yeah and then you've also got say usage and that shows yeah what percentage of your and then a lot of the settings yeah does that also show how how much you're going to get in your rolling window of use?
No. So what I've done is I've also got a bit of money sitting in my account just should I run over the limit. Yeah, I do too, yeah. I think that's like $25 there or something. Yeah, yeah, yeah. But I've not had to use that since I upgraded to the larger plan. I've gotten close, but not quite. Okay, another nerd question. Do you VPN into the U.S. and set up your Anthropic account as though you're a U.S. customer and pay in U.S. dollars, which is considerably less than buying it in pounds or euros? No, I think they just charged me in U.S. dollars. When I've looked at it in euros locally, it's considerably more expensive because in Europe, I'm paying VAT, which is 24% in Greece.
And so it adds a lot of that extra money to it that it doesn't if I'm a U.S. customer. Now, in my case, my company is still U.S.-based because most of my customers are there. And setting up a company over here is a whole different thing, not quite as a full citizen yet. Anyway, another little tip and trick people might think about. This is amazing. This is so much to process. This is the reason I love, love, love Talkadoopian because we could. And on a podcast, I get to ask you all the questions that you wouldn't necessarily have time for. Yeah. I'll ask you one more idea and then I'm going to be done. Yeah. In my little training classes that I do, which are fun, what do you think about having like an AI dev jam session kind of right after the class?
I've noticed that a lot of people stick around after the hour. And I think it would be really fun to basically dovetail it into, after I present, here's how you do these things, to basically say, hey, here's some more detailed stuff I'm working on. Everybody else, what do you have to bear? What's the discussion? What's changed? What new models do you have? What new tools do you have? And just basically have it be a share. Do you think that would be useful? Yeah, that sounds pretty cool. I'll try and experiment then. Yeah, that sounds... You come to a lot of the classes. And I know, I mean, for my benefit, I would love to just ask you more about a lot of this stuff and give you a chance to present this stuff, which would be another really great benefit to the community.
And quite a few people are coming to those classes now. Yeah, I mean, I'd be happy to discuss this. Yeah, no problem. Yeah, okay. Because it's, like I say, it's very much a learning thing on that. That's very early. That's very early days. But the fact that it's actually, you know, got it working that quickly. Well, I say that quickly. I mean, we're probably looking two, three days to actually get it to that point. Yeah. Yeah, true. And it sounds silly to say, it sounds silly to say that because it's doing stuff that, you know, five years ago, 10 years ago would have been magic. I know. I heard some stat in a podcast today that there's a function, I can't remember which one it was, but that it was 90,000 times faster than having a human do the same task, basically.
Which is kind of insane to think about. But yeah, you can also, the other thing is that Clod code in the terminal, it supports hooks. So what you can do is you can have hooks that run before or after certain events. So what that means is you can, you don't need to just give it the prompt at the beginning, but you can have kind of an additional prompt reminder before it does any particular type of event. So when I was building the app, one of the things I found was that I kept forgetting to update the swagger schema so i go and add i change an endpoint and then i'd realize i forgot to change the swagger schema i have to go back edit the swagger schema recompile recompile it have to basically send it off to apple to get approved by apple again download it from apple all this hassle so i basically now i've got it one of the things it does is it checks if i've touched any of these files i now need to go and rebuild the swagger schema um automatically so it's it's taken away
basically systematizing stuff takes away the opportunity for me to make errors that's a big statement yeah well there's plenty of there's i'm more than capable there's plenty I'm still making, but the thing is, as I'm recognizing where I'm going wrong multiple times, I can actually now do something about it and improve the process. Amazing. Well, let's stop the recording here. I'm definitely looking forward to seeing more. Everybody can be watching this space when the tool comes out. Take a look at the show notes for details on that. Ian, thanks so much for your time. Well, thanks for having me.
Have something to share?
Join me for a conversation on ClarisTalk AI.