# Leveraging Local AI for Content Workflow Automation

**Podcast:** All Things Product with Teresa and Petra
**Published:** 2026-02-10

## Transcript

Hi, folks.
This is All Things Product with Petra Wille.
And Teresa, of course.
And we're so happy you're here.
Teresa, I started to use Claude Code four weeks ago.
Yes.
I finally found the time.
I would have loved to find the time earlier, but I couldn't.
Life was lifing, so I couldn't.
But now I found the time to do so.
I have questions.
Okay, let's get into it.
First of all, if listeners don't know.
Starting in November, I actually late October, I have this monster series on Cloud Code and how it's completely changed the way that I work and how everybody should use it.
And I'm not being paid by Anthropic.
I'm just genuinely a super fan.
I saw your LinkedIn post yesterday that you've not sponsored.
So I'm thrilled, Petra, to hear that you're trying it.
What are your questions?
You're pouring yourself a drink.
This is about to get serious.
It's basically hot water.
This is how rock and roll it is these days.
Okay, what I did, I did the basic setup, of course.
I already...
Wait, can I ask you a question?
Yeah.
What inspired you to do this?
So I'm an extensive user of the project setup.
both in ChatGPT and Claude Webb.
And so I have well-trained projects for various things in my business.
For example, I have one in ChatGPT and Claude for writing my newsletter because it's a lot of repetitive work.
So for example, I collect interesting reads, blog posts, articles, talks throughout the three months period.
And then at some point, I only have these titles and links.
And then I go say to the AI, can you go fetch abstracts, write me a quick summary.
I look through it.
I usually have comments what I liked about the post and then write me this little abstracts.
And then I edit them.
Then my editor Melissa edits them.
And at some point, they make it into the newsletter.
And there are many, many things.
Sometimes I'm repurposing some of my book content, whatever it is, right?
So I have a project specifically.
for the newsletter.
And for me, it's mainly the content creation bit because there were limitations to these setups and the limitations were I wrote books and I wrote a lot of other content pieces and I have tons of podcast transcripts and you cannot preload chat gpt web and cloud web with all these documents that i ever that i ever created and have on my hard drive and i found this idea really compelling that cloud code is installed on my machine and is on working on the files on my machine um so there was another reason why i thought like this is interesting and then the other reason is i'm a geek and i love to play with these things so Another reason for give it a go.
That was my, I think my three main motivations.
So the question, is it helpful in content creation?
Are there other helpful use cases that I stumbled upon while using it?
I already did, by the way.
And on my hard drive.
Yeah.
And curiosity.
Okay.
And how did it go?
How's it going?
It's going okay.
It's different than using the tool in the web because sometimes I have this, it's like with a dog.
So I throw a stick into the wood and sometimes it comes back with a helpful answer.
And sometimes the dog gets lost in the wood.
And then you have to really say like, hey, Cloud Code, can you just come back from whatever you do?
So I think you really can feel that it assumes it is working on a code repository still.
So you have to do a bit of work to make sure you understand it's not working on a code repository right now and that your use cases are different.
So that was the first stuff that I realized.
And then you quickly stumble upon this, okay, I can have all these text files with briefings for the AI.
I already had a lot of briefings of writing style briefings that was already existing.
You know the drill, right?
So that's the stuff that we already had to write, the context engineering that we already had to do for our web context settings with Chargivity and Cloud.
So I used a lot of that.
It was already there.
But I needed way more for Cloud Code to work on my machine.
So what is the init state?
So the initialization in the beginning when I start Cloud Code.
So what is the context that I want it to load every time we start talking?
So setting that up took me half a day, I'd say, to a level where I thought like, yeah, okay, now it does make sense.
And now I get helpful answers.
And then the first use case that actually...
Claude Code pushed me into to some extent, I know that the AI is not doing something like that, but you get the idea, was really task management, what you talked a lot about that you're using it for.
For me, that's not a natural thing because I'm usually super on top of my to-do list.
So it's living rent-free in my head.
And in the morning, I do a quick, okay, what are the two things that I really want to get off my plate?
today, but I have like the 38 things that I need to do anytime soon in my head.
It's not, not a problem for me to stay on top of my tasks usually, but you don't use any external task management tool.
No.
Okay.
That is blowing my mind.
Yeah.
I post it maybe, but I'm going to pause you because you've already said a lot of things that some of our listeners might not be familiar with.
So first of all, if you don't...
Unpack it for us.
For people who don't know what Cloud Code is, I wrote a blog post called What is Cloud Code and why you should use it, even if you're not an engineer.
So we'll put a link to that in the show notes.
And then Petra is referring to this slash init that happens the first time you use it.
If you're familiar with Cloud Code, all that does is it helps you set up your first CloudMD file.
If you're not familiar with Cloud Code, those are just rules that you get to write to tell Cloud how you want to work with Cloud.
And here's the secret, clod and clod code is the exact same clod as clod in the browser.
It's just a different context.
And so you get to shape some of that context.
And I also have a blog post about how to create that context and how to set up your clod MD files as part of that slash init command.
And we'll add that to the show notes as well.
Petra, okay, first of all, I am blown away that you don't use a task manager because that's...
My task manager is like infinite.
If I don't get it out of my head, I can't work.
Yeah, I know so many people are overly obsessing with their task manager.
I don't obsess about it.
For me, it has to get out of my head.
Okay, no.
It has to be captured somewhere.
But okay, so you don't use a task manager, but using Cloud Code, this was the first thing you used it for?
How did this happen?
I think it forced me to use it for that.
No, the thing is...
I thought about, okay, what is something easy that I can, talking evals basically.
So what is something easy that I could try where the replies is something that I could easily validate for being true or being false, for things working, for it using the right reasoning.
And I thought about that for a while.
And then really what came to mind was, okay, maybe I can do something like a weekly to-do list first because I have it on top of my head and I'm really good in this kind of thing.
So let's see if Cloud Code can keep up with my brain here when it comes to that.
Because I needed to understand what capabilities...
it has, if it's different, if I need to prompt it slightly different than the web version and all these kinds of things.
So I think that's why I started to do it.
Yeah.
And now that's what it does in the morning.
It says like, after I run in it, if that's the morning, basically.
So, and then it says like, good morning, Petra.
So what, what is on top of mind today?
What do we want to work on?
And then it asks me a few things because it now knows about my projects.
It knows about my priorities.
All these kinds of things, right?
So that's the first thing that I built.
And how did you do this?
Did you just like do a brain dump into a text file?
No, no.
So I was not yet ready to connect it to my Google Calendar.
This is suspicious Petra.
So what I created instead is an export.
from my Google Calendar with limited visibility for the next upcoming four weeks.
I forgot about this.
You use your calendar as your task manager.
To some extent.
We talked about this in a past episode.
To some extent.
Yeah, see, this stresses me out.
I don't want anything on my calendar.
I'd rather have it in my messy task manager.
People, they so do.
Ah, okay.
So I also don't give Claude access to my calendar.
Claude gets access to like...
this walled garden of things that I wanted to have access to only.
So it's really, it's...
Henrik Nieberg was talking about that at our conference this year.
And he said, like, really treat it as an intern.
So you don't hand the intern all the data of your company right away.
So really think of it like an intern who is only with you for the next eight weeks.
And basically what you would tell the intern, that's what you can share with Cloud Code.
And for me, this was a helpful paradigm.
Thanks, Henrik.
So I use that a lot to consider what it's allowed to see.
Okay, so you exported four weeks of your calendar.
I explored four weeks of my calendar.
And then, of course, on my hard drives lives a setup with client folders.
And in each client folder now is a new MD file explaining high level what I'm doing.
So I created a lot of text files and MD files to provide cloud code with a bit more information about to-dos for a particular client or things like that.
So that's what I did manually.
It was just the fastest way.
Because the prompting, and that's one of the questions, it takes ages for it to...
Can I speed up the thinking process?
It's so slow compared to...
Okay, so there's a few things.
Like you can change the model that it uses.
And right now, like as of...
Okay, I know this episode is going to be released in January, but here's a secret.
We're actually recording in December.
And like a week ago...
I know I just ruined the like secret veil of podcast recording.
A week ago, Claude Anthropic released Opus 4.5.
Yeah.
And as part of the Opus 4.5 release, they are defaulting the model to Opus 4.5 and Claude code.
Ah.
So you can change the model and you can change it by just typing slash model and you can change it back to sonnet.
And actually for a lot of tasks, you can change it to haiku.
And haiku is super fast.
So if you're just using it to tell you like what's on your calendar and what the next task is and what the next to do is, haiku is more than adequate at that and is super fast.
Okay, that's definitely something I need to play with.
Yeah, and then the other thing to keep in mind is when you're in Cloud Code and you have your little prompt line on the right side underneath the line, it'll tell you if you're in thinking mode.
Yeah.
You don't need to be in thinking mode for the vast majority of things.
Okay.
And so both those things, like turning off thinking mode and then playing with the model based on the task required.
So like I use Cloud Code.
I have a little bit more robust task management because I can't work if I have like 14 things in my head I'm trying to remember at the same time.
So like I use task management as like a get it out of my head so I can use my whole brain to write this blog post.
and not remember things.
So I have like, I basically moved my task management to be markdown files.
By the way, I'm going to release this as a plugin for people who are interested, and it may be out by the time this episode releases, in which case we'll put it in the show notes.
But every morning, Claude looks through my tasks and aggregates everything that has a due date for today.
So it's like creating, it's almost like when you're creating.
I already saw it live working.
It's wild.
Yeah.
It's like if you, if you created a to-do list on paper, that's kind of what Claude is creating for me.
Yeah.
And then the part I really love about it is like I used to use Trello and let's say I'm doing a job, I'm working and I have a thought in my head like, oh yeah, you got to call that person.
I have to stop what I'm doing.
Go to Trello.
Go to my web browser.
open Trello, which is stupidly slow, click new card, type in what I want to remember, click a little date picker to set the date of when I need to do it by, like save it, make sure it ends up in the right list.
Like it's a lot of work.
I know this sounds ridiculous now, but let me tell you what I do now.
Now, Claude is open 100% of the time in a little terminal window on whatever screen I'm working in.
So all I have to do is like, move my mouse to that screen, and I type new task, call Petra, due tomorrow.
I'm no longer typing.
I'm using whisper flow.
Yeah, I can't get into it.
I tried it.
After our episode, I tried it.
I hate the way it tries to guess what I mean.
It turns out I mean what I say.
I just want straight dictation.
Yeah, I can relate.
I can type new task, call Petra, due tomorrow in five seconds.
Right?
And then I go right back to my work.
And I know that Claude is creating the right task.
I know it's going to show up on my to-do list on the right day.
And then I've extended this.
I have tasks that's like a thing that has to be done on a certain day.
I have ideas, which are like, I'm going to do this someday.
And it could be in progress or not.
I have memories, which is just like, memories are like, I need to remember this thing.
And so what I used to do is I would always put it on my task list because that's the thing that has my attention, but it's not really a task.
And so it would just stay there forever and it would clutter my to-do list.
That's kind of similar to my link collection for the newsletter, right?
So that would be kind of...
Kind of, but this is like random stuff.
Like in Ghost, my blog platform, I always have to add this like table hack to clean up tables and I never remember it.
That's a memory.
I can't ever remember the teleprompter software I use.
That's a memory, right?
Like it's just this little stuff.
And I used to literally put it on my Trello board because like I look at my Trello board and that's where I'll find it.
But that's the wrong place.
Like it's not a to-do list.
And so then my Trello board would be a nightmare.
And so like, and same with bugs, right?
Like you hear about a bug and you want to file it somewhere.
For me, it would go on my stupid Trello list.
And now like, now I can design.
Exactly the workflow that matches my brain.
And so I have.
And so I have.
And so I have.
Yeah.
Yeah.
For me, it's still.
So I'm currently writing my January episode of the newsletter.
Okay.
Took me twice as long this time than it usually does because of cloud code.
Because you're building new workflows?
Yeah, exactly.
Because I at least try to build new workflows.
Okay, tell me what you're building.
Like, what are you doing?
How are you using Cloud Code to help you with your newsletter?
So, for example, the main topic of the next episode of the newsletter, which will most likely be out by the time we're speaking or will be then published a few days after.
We'll talk about these annual reviews that most product leaders have to do.
That's the feedback sessions that HR requests you to do.
You have to put in some information into some forms and then file it basically.
And so many coaches are currently complaining about preparing those and yada, yada, yada.
So I thought like, oh, this is an important topic to talk about because I have a ton of things to say how they could be more effective and how could...
They could be less time consuming on your calendar.
So that's the piece that I'm currently writing.
And the first stuff that I use cloud for is a cloud code is asking it.
So where else have I been talking about providing people with feedback or running such a coaching session, good questions for coaching, whatever.
Right.
So that was the first stuff.
And you gave it access to like your blog.
I gave it access to articles and your blog articles, my books.
all the things that I already have sitting on my hard drive, basically, all my content creation stuff.
And that's what it was really good with.
So that was an easy task for it.
The list that came up was really nice, extensive.
I would say it was all the spots that it could find.
It was way better than I, because I was, I vaguely reminded some spaces where I talked about it.
But for example, It reminded me that I actually have published on my blog, my chapter on how to give feedback that is usually in my book, Strong Product People.
But there was a free chapter that I gave away at some point in time.
And I didn't even remember.
It was like, hey, you can reference that in your newsletter because it's published anyways.
That's a use case that I feel like Claude saves a ton of time.
Yeah.
Because it can just gret through folders and it's going to find those things.
Yeah.
And I think it will be doing the exact same thing with user interviews.
So when was the last time we conducted three user interviews on this particular topic?
It will go find it.
When was the last time a client mentioned this particular thing in one of our user interviews?
I bet it will find it, right?
Who is the person who said this funny thing?
Yeah, exactly.
Or we could do a similar thing with the product at heart.
videos, for example, and talks, and you can extract some of the information from there.
So that was solid.
And then, but for example, with my links, that sometimes was this kind of throwing a stick into wood and the dog was not retrieving.
When I had my collection of links and I set up both ChatGPT project and CloudWeb project in a way that it actually comes up with pretty helpful Hey, that's the link.
That's the website it's on.
That's the title of it on the website.
That's the abstract of it on the website.
You had the following comment in your document.
I call it in a calendar entry, of course.
So you're putting in a, you're saving your links on your calendar?
Yeah, I do.
Like on the day the newsletter goes out or what?
No, on the day I start writing the newsletter.
Oh my gosh, Petra.
I am, I like, I'm terrified to see your calendar.
It like makes me anxious.
So as you read things and you find things you want to add to your newsletter, you add them to your calendar on the day you're going to create the newsletter.
Yeah.
Okay.
So then Claude can see this because you exported your calendar for Claude to see.
And what are you asking Claude to do to like, go look at all those links and pull out like metadata, like a title and an abstract.
Yeah.
And the exact prompt that I'm using in Claude web.
Did not work on cloud code.
And I think it has to do with what you said.
Maybe it's model or it's mode.
What's the prompt?
Roughly.
What's the task you're giving it?
So the prompt is rather precise.
Whisper flow, right?
You can just like talk into the computer.
So it's rather precise.
It's exactly that.
Go look at the website, the link where the link is pointing to.
Okay.
Get me the title, the author, the abstract if you can find one.
And then take my note for this link and put it underneath and then return the list.
Okay.
And Claude in the web did this really well.
Claude in the web does this really well.
Okay.
And Claude code spiraling, looking at the web, looking at the web, looking at the web, trying to retrieve information, trying to retrieve information for pages and pages and patients.
And then at some point I need to say, stop it.
Not like looking at things other than your links.
I don't know what it does.
It's not that talkative, but it always just like says web search, web search, web search, web search.
Oh, okay.
So it's stranded somewhere.
I would need to look at it to be sure, but what it sounds like, it doesn't realize that you have the URL in the calendar of it.
No, but I even tried to paste it into the conversation.
Okay.
And it still is sometimes, sometimes it works.
And at some point I started to experiment with shorter lists.
I thought like, maybe I start with four links and let's see if it works.
Then it does.
Are you doing this all by just having a conversation with Claude in Claude code?
I try to create text files and say like, can you go look, take the content of the text file and go execute?
That didn't work either.
Oh, I'm really curious what's going wrong here.
Because Claude can, so Claude has a tool called WebFetch where it can take a URL and grab the contents of the page.
It will run into problems.
Like some people block LLM crawlers, but you're going to have that same problem on Claude Web.
I suspect there's something it's misinterpreting about your instructions.
Yeah.
Because it sounds like it's trying to find those links rather than using the links you're giving it.
Because I've really tried the exact same prompts in CloudWeb because of, maybe it's me.
But it's not the exact same prompt because your CloudMD files are now part of the context, right?
So there's something in there that's causing it.
Here's how I would debug this.
I would just stop Cloud and be like, Cloud, what are you doing?
I'm serious.
I do this all the time.
Like when Cloud is doing the wrong thing, I just go, Cloud, what are you doing?
And it'll be like, I'm trying to find this and this.
And I can be like, you have the link, just go grab the link.
And then it'll be like, oh, okay, I'll just go grab the link.
And then I'll be like, okay, Claude, why was this confusing to you?
Like your prompt says, go grab the link.
And it will tell you, oh, because in my system prompt, it says this, and it will tell you something from your Claude MD that's confusing it.
I haven't even considered to ask it what it's doing.
Yeah, I asked Claude what it's doing.
I'm like, Claude, what are you doing?
Read the readme.
It's in the readme.
Stop what you're doing.
You're doing it wrong.
Yeah.
It would be so nice if the Whisperflow would even record my tone of voice, but it doesn't.
So even if you could like irony in brackets or something like that for Claude Cooper, it doesn't.
So that would be, maybe if some product manager from that company is listening, I have suggestions.
Okay.
So links didn't work well.
That I think you can fix.
I have Claude do a lot of link fetching for me.
Maybe we could do a little troubleshooting session aside from our podcast recording.
I think you can get that working.
What else?
Yeah.
The thing is that I'm not frustrated yet because I think it's perfectly fine and all under understanding the tool and everything like that helps you understand the tool, right?
So it's not a problem.
Try asking Claude what it's doing.
And if that doesn't help you fix it, let me know.
Yeah, I will.
Yeah.
And then...
the content that it's currently creating is not as good as in because i don't know why this is the case and i have to debug that as well because what i did is it has all my previous writing and it has a tone of voice briefing it has a writing style guide briefing so if i like longer paragraphs or shorter ones or short sentences or longs and all these kind of stuff it already exists and it's basically um fed into Claude code.
How are you telling it where those files are?
That's a mix.
So some of that context is in Claude MD initialization, and then I have other ones per, so there's layers of these instructions.
Yeah, so the thing you want to make sure is that where you're working with Claude has the right layers.
Yeah, yeah, exactly.
And you can ask Claude this.
You can be like, hey Claude, what do you see about my writing style in your context window?
What Teresa means, audience, is you can start Claude code on basically any folder on your hard drive.
But not just that.
Not just that.
Like you can start it in any folder and it will read a local Claude MD file.
Yes.
And then it will read your global Claude MD file.
But the thing is, it also cascades all the way down.
So it's really easy to make a mistake thinking that Claude has access to something and it doesn't.
Yeah.
And that's what I'm currently debugging.
So there was today's work that I tried to understand, okay, where are missing pieces of information that makes the content creation?
So it still does solid writing, but it's not my tone and it's not my style yet.
And I need to figure out where their information is basically.
My guess is you have context in your project that isn't getting into cloud code for whatever reason.
And so what I would do is like when cloud does something that doesn't feel right, if you have something written in a cloud MD that you think is steering it and it's not listening to it.
Just ask Claude, like, hey, do you see this sentence in your context window?
And then the other thing to be mindful of is you can have too much stuff in your Claude MDs.
So you don't want, like, especially your global Claude MD, you want to keep that as lightweight as possible.
Yeah, that's currently rather lightweight.
Because for your context, you want to give it everything it needs to do the task well, but no more.
Anything that's unrelated to the task will be a distraction.
Yeah, it should be somewhere else.
Yeah, that's what I already learned.
Yeah, but it's fascinating.
It's interesting.
I will continue using it.
I will continue setting it up because newsletter is only one small fraction of what I do, right?
I think it will not...
be what it will be helpful with with the course creation is again the research part so where i've been talking about these already um even my illustrations it can read my handwriting so far that worked well so i can even go through my illustration and suggest hey you already drew something like that you could use it um so i think this content retrieval is quite a cool use case for everybody who has written a lot or has a lot of written content again user interviews um sales, market interviews, trade fair interviews, all these kinds of things, right?
So that could be easy to buck reports, customer service tickets.
So I think that is all amazing.
And you could easily train it on that and use it for that.
Will save you a ton of time.
Yeah.
And let's see what are the use cases I will be finding in the future.
So I like a really similar use case to what you're trying to do with your newsletter is I use clod code to write my...
So for my podcast, Just Now Possible, clod code generates the episode title, the episode description, the show notes, the chapters, all based on the interview transcript.
Nice.
And I do that with one slash command.
I literally type slash JNP dash metadata, and I give it a transcript file, and it produces another file for that episode with all that just done.
I almost never have to edit it.
Wow.
But I've written like a really detailed style guide for what I prefer on how those are generated.
Yeah.
Right?
And I have a different style guide that I use it to help critique my blogs.
I still write my blog posts, but I use Claude to act as an editor and critique them.
I do this mainly for shitstorm prevention, I have to admit.
And I...
But I don't co-mingle those style guides, right?
So I have a style guide specifically for podcast metadata.
And so like I'm trying to isolate what does Claude get given the task.
Yeah.
So that it performs better.
But you can, I think you will find the more you play with it, especially for this like throwaway writing, right?
Like it's not...
It's not your opinion piece.
It's just I have to write a couple sentences here to do a transition from here to here.
I think Claude is more than capable of doing that writing for you.
In your voice.
Yeah, in my voice.
And I think it will be good as a sparring partner at some point for...
I can feed it with all...
the LinkedIn analytics, for example, and what are people really interested in engaging with.
And then I can have, because I always have a long list of topics that I want to start working on and want to write about.
And that could be easily a task for Claude Code, I think, to say like, hey, what is the most promising stuff on my list that my audience would be interested in reading?
help um use it for a bit of prioritization help and these kind of things so stuff that i'm usually overlooking because right now frankly speaking a lot of my content creation is stuff that i want to do yeah and I could easily prioritize other things on my want to do list that is currently a bit more on the bottom just because if we would see okay audience is currently more engaging with this other thing we see it for example for our podcast on all the AI episodes so we see people are massively interested in that right That's an obvious pattern, but they're less obvious pattern in my Petra Villa LinkedIn audience analytics, for example.
Yeah.
So I do a lot of LinkedIn analysis with, I have this process.
I got it from Andy Crestedina where I export my, Andy Crestedina is from Orbit Media.
He's like a content marketing guy.
I export my analytics from LinkedIn every month.
And then I use ChatGPT to generate heat maps.
I could use Cloud for this too.
My admin actually does this for me now.
And at the time we were mostly using ChatGPT.
I think she still uses ChatGPT for it.
But it generates heat maps based on topic and category.
So we have different categories of posts and then different topics.
I will tell you, it is really hard to find trends.
Like I've been doing this for two years now, very rigorously.
The only trend I feel confident in is the posts that perform the best is my stupid off-the-cuff, like, I'm writing a random post from my kitchen on my phone about whatever it is I did that day.
That's what performs the best.
Everything else, and we've tried it.
Everything.
But what we figured, for example, for a product at heart with the help of VI, is...
Because everybody currently advises you, everybody knowing how social media sources are made, basically, is saying like, don't do a visual for every post.
Text posts perform better.
So often heard that advice.
For Product at Heart, we really, it's by the way, the case for my personal account as well.
So if there's only text, it has more engagement and more people looking at it than if I have a nice visual.
And I hate it because I love to brighten up people's days with my hand drawings or whatever random visual.
But the posts don't perform that well.
But for Product at Heart, it's different.
For Product at Heart, it's still, if there's visual to the post, 10 times more engagement.
Wild.
I do think like company accounts versus individual accounts perform differently.
Yeah, maybe.
But we would not have figured that out because nobody does the peanut counting themselves manually.
And I'll say a lot of the rules people say are true are not true, right?
You have to.
You got to just run tests.
Exactly.
The good old product work people don't believe what other people say.
Bring your own data.
I'm excited for you on this journey.
I will tell you, like I'm now like for writing, I've now built a whole suite of tools.
Like I've designed a headline brainstorming tool that matches like my preferences on headlines.
I have an SEO tool that literally like integrates.
And when you say tools, what are we talking about?
Yeah.
So in Cloud Code.
There's a lot of ways you can create tools.
You have slash commands, you have sub-agents, you have hooks.
I'm going to be writing a blog post about this.
We'll see if it's out by the time the episode comes out.
There's a lot of tools you can use to create reusable prompts, workflows, even run code.
And then you can bundle these into a plugin.
So when I talk about I'm going to release plugins around some of this stuff, it'll be a collection of tools.
So in my writing workflow, I have a headline tool.
That is a slash command.
I have a, so I type in slash headline and it will run through, like, I think I get like 25 headline suggestions.
And then we iterate on what I like and what I don't like until we get to a final headline.
And then I have an SEO command.
And what my SEO command does is it reads the article.
I run it after I've written the article.
It reads the article.
It starts to generate.
both short tail and long tail keywords that it might potentially rank for.
It uses an API to run keyword volume queries and then gives me a report of this is the SEO potential of this article.
And then that helps me with like, should I change a word?
Like, should I change a term to be a different but semantically equivalent term so it can rank a little bit better?
I never did any of this stuff before.
Like, I just wrote what I thought.
And I still start that way.
I still write the article based on what I want to write about.
But this allows me to like fine tune and tweak it a little bit just to get a little more performance out of it.
I run a sub agent that's a fact checker.
And depending on the article, it fact checks in different ways.
So if I'm doing an article where I interviewed a bunch of people, the fact checker looks at every single claim in the article.
that's related to an interview, and it has to pull evidence from the transcript to support that claim.
And if it can't find evidence of that claim, it gets flagged for me to review.
Okay, then you double check.
Okay.
Right?
If I'm citing research in the article, it looks and makes sure that it's a real paper, that I got the author right, I got the year right.
It looks at that paper and makes sure that what I'm claiming is true to that paper.
Right?
Um, so I've built these like little tools that I use at different points in my writing process.
Nice.
Um, I probably won't.
Yeah.
It's just fun.
And I don't, it means that like, I do way more rigorous stuff than I've ever had time before to do.
Yeah.
Yeah.
I, I, I see that as well coming up on the horizon for my work.
I have to admit.
Yeah.
Let's see.
Let's see what it helps us with.
It is.
And right now it's.
time consuming.
Yeah.
I bet you are already past that point where it takes more of your time than it helps you with.
So you said like there's more rigorous stuff that you would usually not be doing without it.
Yeah, it's still both because I feel like I'm always pushing the boundary of what I want to do more with, right?
So the return on invest, let's see.
For me, here's what I'll tell you.
I used to write 8,000 words a month.
I now write about 35,000 words a month.
And I'll let our listeners be the judge.
Go look at my recent product talk posts.
Has my quality degraded?
I can say that pretty confidently because I've gotten some of the best feedback on my writing in the last few months.
Yeah.
Like I don't think my quality is degraded, but my output has gone way up.
Yeah.
And for example, my newsletter, I think would be more substantial because of things like, hey, you published this for free already.
Then why not sharing it with the people that it's available for free?
And otherwise I would not have even thought about it.
Okay.
I know we're already really long on this episode, but I just thought of another use case that I recently started with that I absolutely love.
I will share it as a ton of work, but I absolutely love it.
So I think I mentioned in a previous episode that I'm starting to dig into the research on synthetic users.
So I'm literally like doing a broad literature review of everything related to synthetic users and digital twins.
I decided as part of this project, I was going to start a Zettelkasten.
I'm saying that wrong.
It's a German word.
Yeah, Zettelkasten.
Do you know this?
Slip boxes?
Yeah.
Okay.
I've known about this concept for a very long time.
It's always intrigued me, but I've never bothered to do it because it's a ton of work.
Here's the idea behind it.
It's like a way of taking notes, kind of.
People often describe it as note-taking, but as I get into it, I'm learning that's not really what it is.
It's a way of collecting beliefs and examining your beliefs.
And this is the part that I'm really enamored with.
It's like critical thinking embodied.
So here's how it works.
You have a main note or a permanent note.
There's all sorts of different similar language, but you write a note.
Like as you're reading something, you pull out a claim.
So a claim might be synthetic users can simulate human behavior.
That's a claim.
Okay, well now you have to like, what's the evidence for that note?
And what's the source?
And what are your thoughts about that source?
And is that strong evidence?
And then you might have like, well, I see a limitation.
That limitation can be another note and they link to each other, right?
And so you're building out like independent claims that all interlink with each other.
Okay, this is really hard.
It's really hard.
It's really hard to have one claim per note.
Like it's just like this rigorous process, right?
But here's the deal.
I want to write a research report on synthetic users.
This is a very controversial topic and I want my research report to be rock solid.
So I'm putting in the work to like examine literally every single claim that I make.
But I don't know that I would have known how to do this because like, You read the average blog post on Zettelkastens and you can tell me that I'm butchering that German word.
But it's fine.
It's fine.
It's not like the average person doesn't do this rigorously.
So like when I read the random blog post, it doesn't really tell me the right way to do it.
Well, it turns out Claude is very familiar with this concept.
Claude can read all my notes and give me feedback on do I have multiple claims in a note?
Am I missing important links?
And so this like really rigorous thing that was like a little bit tedious.
a lot a bit tedious like claude does a lot of the tedious linking and like it doesn't even bother problems and i get to just focus on like what do i think of this claim and that is not something i would have ever done without the aid of claude like there's no way i would have tackled this that's super nerdy theresa oh my god it's so nerdy but i'm loving it yeah that's what we love you for yeah It's really funny because I consider myself usually pretty nerdy already, but you are Petra times 10.
Well, this is like an intellectual adventure.
I think adventure is the right word.
I am going in to study a topic where I already come in with my own biases, but I don't want my own biases to color what I learn, and I want to be open.
to changing my mind.
Okay.
If I, like I'm a nerd about critical thinking.
I mean, I read John Dewey's How We Think, which is a 200 word book that took me 15 hours to read because it's super dense.
And like, he's a 1900s philosopher and he writes like a 1900s philosopher.
But to me, like this is one of the most important things for us to study because it's like critical thinking is the heart of what we know.
It's such a future skill.
Yeah.
Right.
So for me, What appeals to me about this process is John Dewey says, if you hold a belief, it is your responsibility as a critical thinker to examine the priors that belief depends upon.
Right?
Well, that's what this enables you to do.
It allows you, you're literally identifying a belief and then the downstream effects of that belief and the upstream effects of that belief and linking them and looking at each one in isolation and like really questioning, do I believe this?
OK, so to do this for a topic that's like generally controversial that I want to be able to write about and write about confidently, like this is awesome.
Yeah, that's what, for example, I say with my I use it a lot for preventing shit storms kind of thing.
I know it's a catchy headline, but it has it goes along similar lines.
So I'm claiming a lot in my post sometimes.
So is that has that any potential in being really kind?
confrontational or are there any biases that I don't see or all these kind of things.
And I think if you're writing a lot and put a lot of content out there at some point, yeah, a responsible writer would ask themselves similar questions to some extent, right?
Not with the same rigor than you, Teresa.
Here's the thing.
I know that not many people like this stuff.
They don't want to read the academic papers.
It's intellectually effortful.
We don't have time.
It's messy.
It's very ambiguous.
Guess what?
I love this stuff.
Like to me, it's a puzzle I'm unraveling.
I love it.
I want to dig in and be like, this study says that, but that study said, like, I love this stuff.
So what a great skill for me to then do all this work that I love doing and then translate it to be meaningful for people.
They don't love to dive into other studies.
So that they get the benefit of all that research and they don't have to do the work.
That's awesome.
That's like a win-win.
I get to do something I love and other people get to learn a little bit more about what's true.
That's what motivates us.
That's what motivates us.
Teresa, we're wrapping this up.
This is our longest episode ever.
I hope people lasted to the very end.
Thank you for helping me here live clinic with Teresa Torres on Petra setting up cloud code.
This might be a two commute episode.
Yeah, morning and afternoon.
Sorry, folks.
Exactly.
I hope you spent the whole day with us.
Thank you, Teresa.
Thanks, Petra.