How Ahrefs team is automating blog content with AI | Ryan Law (Ahrefs)

So Ryan, I was thinking today, uh, it didn't take us too long to go from this AI thing cannot really do great content to, oh my god, this is amazing, right?

And here we are. Uh, you have devised an AI content automation workflow that you use to actually publish quite a few articles on HF's blog already. And these

are good articles. These are great articles. And yeah, full disclosure, I

articles. And yeah, full disclosure, I haven't seen it yet, so I'll be checking it out live together with uh anyone watching it, and I'm excited. Yeah. Do

you want to say a few quick words on what people are about to see? Yeah.

Well, exactly that. So, we've obviously been tinkering with using AI in our content workflows for years at this point, and it's always been very effortful. It can be helpful, but you

effortful. It can be helpful, but you have to sink a ton of time and energy into it. There's still a lot of manual

into it. There's still a lot of manual stuff that has to happen. I kind of feel like that's not the case anymore. Uh

it's a bit spooky actually. I think

since Claude code is probably the big thing that has changed this this kind of agentic workflow where Claude can make some decisions on your behalf and you can provide it with some guard rails to

actually make it do things in a certain way. Um so we've basically yeah built I

way. Um so we've basically yeah built I call it the blog pipeline. Uh and it is a kind of content automation system for new articles and for content updates. Uh

we've done maybe like 30 article updates with it so far. Um published maybe uh 10 15 articles. Got maybe something similar

15 articles. Got maybe something similar in progress at the moment.

Um yeah, it's been pretty cool.

Let's review it. Show it to me.

Yeah, let's do it. Um so obviously two things on the screen right now. We have

a terminal and we have Claude code running in that terminal. So that is just a folder that I've called blog pipeline and claude code is living in that folder and it will do stuff in that

folder for me when I ask it to. Um and

we've got VS Code over here. This is

just a really good way of showing you the contents of that folder in a way that is a bit easier to understand. Uh

these are all the folders on the left hand side and they've all got files inside them. And I of course I asked

inside them. And I of course I asked Claude I said I'm going to present this process on a podcast with Tim. give me

some notes and visualizations to help explain this. So, it added a handy

explain this. So, it added a handy little podcast folder here uh that has some notes and visualizations to make this a bit more interesting.

But the basic premise is uh we have basically set it up such that there are maybe 23 or so uh skill files in here.

You can see them in this folder. Uh and

skill 23 skill files. That's that's a lot.

That is a lot. Um, and each of these skill files is basically a process. It

is a process that at some point during creating content or updating content as a human, we generally do something like this or very similar to this. Um, and

this is just a markdown document with very you lost me already. I I kind of know what skills are, but you lost me already. Let's start from the beginning.

already. Let's start from the beginning.

Where does the process start? What do we start from? Do we start from a keyword?

start from? Do we start from a keyword?

Do we start from an idea? What's the

first step with this?

Yes. Uh so what you can do is so this is a keyword ideas CSV. Um it's obviously a bit hard to see in this format but basically we've even set up a a process

right now where we can use the HF's MCP which is obviously a way for Claude and other LLMs to access HF's data and it will run a content gap analysis for us and I've then set up another process

where it will review this list of keywords and prioritize them. Uh it

looks Ryan, again, let me uh use my Eastern European politeness uh to bring you right to the point. We're not talking about keyword research. We're talking

about creating content. So, let's say we have a keyword or a topic. Uh what do we do with it? We're not we're not talking about keyword research. We want to create content. Let me show you that. Uh

create content. Let me show you that. Uh

let me clear this.

Uh so, I can trigger blog pipeline. And

I can put in a keyword like keyword opportunities.

And if I want, I can add some context to it and explain, you know, if there are points I want to add, I can add that.

And off claude goes. And probably

somewhere from between 8 to 11 minutes from now, it will have a draft ready for review. Um, it goes through, yeah, about

review. Um, it goes through, yeah, about 12 steps at this point, as you can see.

Oh, wow.

It's actually It's actually telling me we've already It's so clever. it's not

letting me do the same one that we've already done before. Um, so there's a research step, there's a reference step where it looks at existing articles on the HF's blog. There's an outlining step

where it turns that into a structured outline. There's a like product

outline. There's a like product annotation stage where we look for opportunities to mention specific HS products. There is a drafting phase, a

products. There is a drafting phase, a citation phase for internal linking and finding supporting sources. There's a

screenshot phase, which doesn't work very well, but we're working on that.

Uh, a preview phase where you can actually preview how it would look on the blog. Uh, and then a formatting for

the blog. Uh, and then a formatting for publish phase where it will add in all the WordPress short codes, all the kind of stuff we need. Um so basically when

you when you ask it to create uh a piece of content around a given keyword you are essentially uh launching a skill

which is a combination of steps where each step is uh a separate skill and basically it has to finish them one by

one or how does it work exactly that? Yeah. So what you can

exactly that? Yeah. So what you can either trigger the skills individually.

So if you just want an outline, you can just ask for an outline by triggering that skill. But I've created these kind

that skill. But I've created these kind of master skills and they are very simple. They exist just to tell Claude

simple. They exist just to tell Claude to work through the other skills in a particular order. Um so this one's

particular order. Um so this one's called blog pipeline and there's also update pipeline. Uh and they literally

update pipeline. Uh and they literally yeah just stitch them together and make Claude systematically work through these processes. Okay. Uh let me go straight

processes. Okay. Uh let me go straight into the uh phase of being critical of this. Uh how is this not a slope? So

this. Uh how is this not a slope? So

what makes this process produce good content and not some generic stuff?

Yeah. So that's a very good question. I

think this definitely works best for one thing on topics that we have already covered in some capacity on the HF's blog. Um, so one we published recently,

blog. Um, so one we published recently, content gap. We have never written an

content gap. We have never written an article about content gap spec or is it keyword gap? One of these two, content

keyword gap? One of these two, content decay and keyword gap. We've written

loads about these concepts generally, but in slightly different contexts, but we've never specifically targeted that keyword. But because this uh is able to

keyword. But because this uh is able to go and look up existing HF's articles and anchor the content generation process in what we've already written,

uh that goes a long way to getting rid of a lot of the problems you'd have. Um

and there are also some topics I think I'm mainly using this for very straightforwardformational topics, things that the LLMs know a lot about.

Um there are opportunities for you to provide some context in it. Uh there's a particular step in here that looks for um opportunities to add information gain. So it actually reads the top

gain. So it actually reads the top ranking articles, summarizes the contents of them and make suggestions for ideas that are not covered but would be useful for the reader to understand

within this.

And I think AI is better at research than a person is as well. Um it can be faster and more systematic about it. It

can go out and look up uh you know the latest research articles, the latest stats, all these kinds of things. Okay.

So, we definitely cannot go through uh all of your skills uh in the course of this podcast episode because there is a lot of uh a lot of content in each of

the skills. But I want you to uh start

the skills. But I want you to uh start from the very first step of creating content and then go to the second and third and highlight maybe one or two

kind of counterintuitive things. So, for

example, what what might people get wrong if they would want to kind of recreate your process? Because uh I'm not sure that we want to just uh give out your process to everyone else if we

want just like open source it and have everyone else have access to the same process. And I would imagine that people

process. And I would imagine that people would want to uh make it uh personal to to them and their voice and their blog and the style of content that they want.

But yeah, walk me through each step and tell me if you uncovered anything interesting about

uh giving instructions to AI on how to kind of improve the output of this specific step if you know what I mean.

Yeah. Yeah. Yeah. So you made a very good point there as well. I think the way we are using this is not as though this is the universal process that everyone the team has to follow. Um,

we've actually set it up such that the team can fork their own versions of this repo. So, they can make their own

repo. So, they can make their own version of this folder and they can modify it how they like. This version

has examples of content that I like and my writing voice uh and it's used as part of the article generation process.

It would be super weird if SQ or Louise uh did the same thing, used my writing voice for their articles. So, it's very easy to actually update it and personalize it. And part of that might

personalize it. And part of that might be changing the steps it goes through your own personal preferences. Um, this

is kind of very unique to me, I think, and that's kind of I think how this should be used. Um, another good point as well, you said you were surprised at how many steps there were in this

process. I think that is actually a very

process. I think that is actually a very very good thing. Um, the more steps you create, the more kind of introspection you have into the process, the better you understand it. uh the more

opportunities you have to actually control and personalize how the content turns out. So one very important thing I

turns out. So one very important thing I learned very quickly obviously I could just set this process in motion and it would give me an article in 8 minutes and either it's good or it's bad. It's

quite hard to work out how to fix and improve the process if you do that. So

actually at every single step of the process um you can actually see it it will give me an output at every stage.

So if something goes wrong, if I don't like the article or how it turned out, I can go back and see which part of the process it didn't work very well. I'm

actually surprised that when you initially tried to launch this process, you said that you would wait like 8 to 12 minutes because I was expecting that

you would actually babysit it from step to step. So you would see the output of

to step. So you would see the output of the first step, see if you want to refine it, if if it's according to your expectations, and then allow it to go to the second step. But it's all just

batched for you. And the thing is that that's actually an interesting tip. This

is exactly what I was looking for, some tips for people of uh what they need to look for when they're building this themselves. And the tip is make sure

themselves. And the tip is make sure that the process saves the output of the step. So if you don't like the final

step. So if you don't like the final thing, you can go step by step and review at which step kind of it went sideways so that you could uh like give

it more instructions or refine the the skill that that refers to the step uh and make it uh do over again and see if that would help. But yeah, I'm surprised

that uh you let it run for like 8 to 12 minutes. Is this the point that you

minutes. Is this the point that you trust it well enough? you like your steps or Yeah, probably it's that I I don't see any other reason why would you would just let it cook for so long and

follow all the steps. Yeah, great point.

Very importantly, this is uh actually months and months of refinement has gone into this thinking and the process in here and actually the last podcast episode we talked about where we had the

custom GPTs, it was like the kind of baby version of this process. So a lot of the skills I have in here are things that we improved and refined and did handhold and babysit as part of that

process. So we'd already written these,

process. So we'd already written these, already tested these, already made dozens and dozens of article outputs with them and kind of learned to refine them. So the thing that Claude does very

them. So the thing that Claude does very well is just stitching those together.

Um actually automating that process.

Okay, let's go step by step. The first

step I think I see it though it is quite small is research, right? any like one or two tips to that you saw that would significantly improve the output of this step.

So it does a combination of things.

Maybe most people would assume you know keyword research is the most important thing to do and we have that in here. It

goes and gets a bunch of HF's data from the MCP uh related keywords parent topic all this kind of thing.

We don't have this yet but I've asked for it. What is more important I think

for it. What is more important I think is going and looking at the existing SER the content that is ranking and analyzing that and seeing the topics that are kind of consensus and commonly

used there opportunities to differentiate from that um that is what AI content helper is perfect for doing but we don't have the endpoint for that yet so this does a kind of laborious

manual version of that but what is ex what is exactly like what are you asking it to do do you ask it like open the top ranking articles for this keyword and

what and and read them and summarize them. What do what do I ask it to do?

them. What do what do I ask it to do?

Like give me something interesting about the research step. Yeah, let me try and find it. Here we go. This is the skill

find it. Here we go. This is the skill file.

Um so it starts with keyword ideas.

It gets uh primary keyword metrics and parent topic. Uh it finds longtail

parent topic. Uh it finds longtail keyword variations that share the same parent topic.

Uh there's some prioritization where it groups them together and discards ones that wouldn't fit the right intent.

Pulls the questions report through the MCP as well. So we get commonly asked questions that people might have related to this topic. Groups them into question themes. So we're not just doing like FAQ

themes. So we're not just doing like FAQ spam.

Uh we get the SER overview. We use that to go and look at the type of content that is ranking, the estimated traffic, all these kinds of things. Uh analyze

the dominant search intent of the SER results. So we can see what type of

results. So we can see what type of content performs best. That's kind of going into this process. Uh and then it looks at the actual top ranking pages.

So it uses web fetch. It retrieves the content. It extracts the headers. It

content. It extracts the headers. It

summarizes them. It looks for themes and gaps in them. And yeah, creates content gaps and opportunities.

Um and you can see an example of the kind of output. So it basically creates a report like a research report at this step. I don't have to see this, but this

step. I don't have to see this, but this is what gets fed into Claude at the next stage of the process. Uh, so you got loads of keyword data, questions to answer, organic results. Uh, you know

what? At this point, as I'm looking at

what? At this point, as I'm looking at how detailed and sophisticated these steps are, I want to say the word overengineered.

M I'm actually wondering if you would like remove half of that, would it just do as good of a job?

Yeah, quite possibly. And that is another really important part of this process. Um I'm always surprised at how

process. Um I'm always surprised at how good increasingly the most like uh frontier most up-to-date models actually are on their own without any input. Um

so a big part of the testing and iteration we've been doing is to we've actually been writing um like test cases. We've been following this these

cases. We've been following this these steps with the skill file and without it and seeing whether the without version is actually good enough. Does the skill

file actually add any benefit to it? Um

a good number of cases the models do a very good job on its own and it just needs a little nudging in a particular direction. So, I expect as we continue

direction. So, I expect as we continue to improve on these, these skill files and these outputs will just get simpler and simpler over time until they're distilled down to the handful of things that are very important for getting the

output that we want cuz yeah, it's probably completely overengineered at this point, I think. But again, my brain wants some kind of structure to what

we're trying to do. So the the structure I would uh so if I were building this process myself from scratch and they needed to start from research of competitors and I know that my

competitors are the pages that are ranking uh at the top. Uh what I would tell uh AI or Claude specifically to do I would tell it to download all the

content uh in the folder. uh and then yeah I would I would ask it to extract from each piece of content kind of the main themes and the main ideas and then

I would ask it to cross reference those main themes and the main ideas between the articles and create me one master document with all of the kind of ideas

stories interesting points uh from all of the content so my my output uh I I don't necessarily need like you had the people also ask questions and that

stuff. I would just ask it to uh analyze

stuff. I would just ask it to uh analyze articles and create kind of a blended master file with everything unique that is pulled from all the articles.

Actually, I do a similar process right now when I when I prepare for podcast interviews uh with uh marketing leaders.

What I do is I do a pretty similar thing. Uh I give Claude uh their

thing. Uh I give Claude uh their previous interviews, links to their previous interviews on YouTube. It

downloads the transcript and then it creates me for each of these transcripts because I don't want to read the whole thing. I want TLDDR too long didn't

thing. I want TLDDR too long didn't read. So I ask it extract the questions

read. So I ask it extract the questions because questions are topics within the interview and then differentiate between uh main questions and follow-up questions where the host is digging uh

more into this topic. So, uh, yeah, you know, like where the main question, where the follow-up question, and then instead of giving me the whole answer, uh, give me TLDDR, just a few sentences

of what the guest replied, give me if there was any hot take, give me if there was any story, give give me if there was any specific number like, oh, we increased our leads by 300% or

something. Uh, and I think there was

something. Uh, and I think there was something else, but I forgot about it.

So, it creates me TLDDR for each of the interviews. And as the next step I ask

interviews. And as the next step I ask it now create me a master TLDDR and this is what I would read uh while preparing for for the podcast interview because it

would give me all the unique information from like a dozen interviews. So it

feels that uh when I want to create a piece of content it's kind of the same.

I want to know what has been said already by uh on this topic. So this is what I would include in the research phase. But uh yeah, you're giving it uh

phase. But uh yeah, you're giving it uh people also ask questions, parent topics, but it it feels that when you say that you're extracting kind of topics from a page, it feels the same

what I'm doing when I'm extracting questions that a host asked uh my guest and then I do do you also ask it to create a master document with everything?

Uh that is yeah basically that research document is the kind of um uh this is the research document it would hand over to the next step of the process.

Okay. Uh we discussed research. What is

the next step?

So the next step uh uh HF's references.

So how how does it work? So this was I actually added this very recently and this has been very helpful. Um

Claude can do a good job writing an article on most topics. It can go and look up other content. That's all well and good. Um, I really wanted a part of

and good. Um, I really wanted a part of the process where, you know, as a human writer, I would go and see what we already have on a topic because I want to make sure a new article is consistent with old things we've written. I want to

interlink between them. Uh, I want to make sure the kind of framing is useful.

I want to be efficient and make sure I'm not repeating myself. I can just pluck elements from existing articles. So this

specifically looks up the target keyword to see what we have already published on that topic, what is already ranking for similar topics and it incorporates

elements of that into the uh like outlining and generation process.

Okay, it feels like it feels again I will try to uh explain it from my perspective. Uh I will try to kind of

perspective. Uh I will try to kind of simplify the process. So it feels the same as research as the first step where you take the pages where you extract kind of unique information from them and

you want to understand kind of the the overall topic coverage uh as pulled from like a dozen different pages and now what you're doing you're referencing our own content. So rather than searching

own content. So rather than searching which are the top 10 ranking pages for the topic, you're going and searching okay what relevant pages does hrefs does

our website already have on this topic and can we pull something interesting from them and again cross reference with my master document if we're saying something unique uh that this master

document is not saying and what's what's important is because our content is very productled and we try to fill our content with use cases of our tools and

data often times the unique bits that uh AI can pull from our content on this topic are those use cases and you can even specifically instruct it so you can tell cloud so specifically look for

whenever we're discussing this topic how are we uh teaching people to use our tools what kind of actionable use cases we're teaching them uh and then it would create you another document with like

okay this is the master document of what all competitors are talking about this topic and these are unique unique insights that I saw published on your blog and here are unique I don't know

use cases uh of your product that I saw in your articles uh on this topic. So is

this more or less what you're looking for? Yeah, exactly that. And this step

for? Yeah, exactly that. And this step is quite simple as well. It's basically

I wanted to provide a almost like a list of modules or sections that could be relevant to this topic that we have already covered so that when it comes to outlining and drafting Claude can go and

look up these examples, incorporate those headers, uh link back to them as an internal linking step just make it kind of an integrated part of how we create content.

Okay. And then we have next step.

Yeah. And then onto the outlining phase.

Um, so let's have a look see if I can find the skill for this one.

So these are this is very similar to what we had in the uh custom GPTs. This

is kind of the editorial process that when a writer puts together an outline, this is how I expect them to do it. Um,

so it's got some very simple core concepts. Um, you know, every uh we must

concepts. Um, you know, every uh we must use the bluff principle. So, every

section must open with the most important idea and then segue to examples, extra context, that kind of thing.

Um, we need to make sure we're logically supporting the thesis. So, the headers must make sense within the context of the title you've created. We need to be exhaustive in how we cover the topic. We

need to be mutually exclusive so we don't have loads of overlap between each of the sections. Um, and again, these are things that if you ask Claude to edit an article and make it me, it does

a fairly good job of that. It has a good comprehension of what that means.

Um, and then you can see, uh, an example out of an outline here.

So, we've got hook, key points, uh, any ideas for transition it wants to include or a specific example it wants to include. It wants to include a table.

include. It wants to include a table.

Um, these are the bones of the article.

you mention a very uh important word.

The word is example.

Uh I can give you uh a quick reference of why I'm talking about it. So I have a bunch of uh skills in my cloud code for

creating LinkedIn posts. Uh

for example, I have uh product based LinkedIn posts when I'm announcing a feature. I have uh podcast announcements

feature. I have uh podcast announcements when I'm announcing that I had a new guest on the podcast. Uh or just regular posts when I have an idea and I want to

kind of deliver it in the best possible and punchy way. The thing is for like each of those is a separate skill that I have created and I have instructed uh uh

clo code of what I'm looking for because when I'm announcing a podcast that's one format when I'm announcing a product update from HFS that's another format.

when I want just to improve a random post that can be about anything that's a different set of instructions. But the

thing is for each of those skills I have a folder where I I have given claude code a bunch of examples here are the examples of my previous podcast

announcements. So that not only you have

announcements. So that not only you have my instructions of how to write them, how to structure them, you have examples of how I did that in my voice already previously and also like those examples

come uh also with engagement metrics. So

it's it even sees which posts perform better, which post performed worse. Uh

same for podcast announcements, same for product announcements, and same for random posts. So it always have uh a

random posts. So it always have uh a folder with examples to reference. And I

almost feel like when it only has instructions versus when it has instructions and like five to 10 examples, I feel it does a

better job when it has kind of the actual examples to fall back to. So when

you're saying that this is there's a step of an outline, I almost want to you to have a folder where you have five examples of outlines of previous posts.

We do have that somewhere. Is it

templates?

Yeah, somewhere we do have that. Maybe

it's in part of the the skill files because exactly the same thing. I you

know, we every time we generate something, we generally want it to maybe sound like us or sound a particular way.

And I used to see a lot of people feed it writing and say, can you distill my writing down to a handful of principles that you can then follow? I was I was always very skeptical of that though like how can you reduce somebody's

unique voice down to a handful of things that then Claude without that example to back it up can actually go away and do.

I think what you you're right the much better thing to do is let the model infer itself from an actual example cuz you know your writing style I don't think is always going to map neatly

across to a five bullet point lists of your writing style or whatever but Claude is a large language model. it can

infer from large samples of text the patterns that do actually exist in your content and that is how it will end up sounding like you. So I totally agree anchoring it with an actual example and saying make it sound like and feel like

this is actually pretty good from what I've seen. Yeah. And this is this is

I've seen. Yeah. And this is this is exactly how people should create those skills in the first place because the way I created my skills is I gave it a bunch of my previous podcast

announcements. I said analyze these

announcements. I said analyze these posts. tell me what I'm doing here, tell

posts. tell me what I'm doing here, tell me what's my style. It would tell me like what it kind of inferred from uh reading my posts and I would correct it if I disagree somewhere. If if it

doesn't feel like it understands what I'm doing, uh sometimes it would understand what I'm doing better than myself, which is funny. I'm like, "Oh, that's that's really what I'm doing. I

just I I was doing it subconsciously. I

didn't understand that." And then for example uh speaking of podcast announcements I would give it some podcast announcements from Lenny Richitzky how he announces his podcast interviews on LinkedIn and I would say

okay analyze what Lenny is doing here.

Uh it would analyze what Lenny is doing again I would correct if I disagree with something and then I would say now create kind of something in between something between my approach and Lenny's approach and tell me what set of

instructions you would come up with. So

basically I'm not creating instructions myself. I don't need to write out

myself. I don't need to write out instructions. I'm giving it examples.

instructions. I'm giving it examples.

I'm telling I'm telling it analyze and tell me what you see like what's the kind of principles behind this piece of content. And then I would correct it. If

content. And then I would correct it. If

I disagree with something, I would monitor what what instructions it is creating for itself and I would correct it. And then like I said, it's very

it. And then like I said, it's very important to have those examples for it to fall back on. Uh because then I just I just feel the output is always better.

Okay. So that's outlining step. Uh like

you said, you you do have some outline examples. Actually, it's as easy as uh

examples. Actually, it's as easy as uh asking Claude, hey, outlining step. Uh

tell me, do we have examples for it? How

are they stored? Are they stored in a text document? Are they stored in a

text document? Are they stored in a folder? And it would tell you. And if

folder? And it would tell you. And if

you don't, you can just say, okay, then create this folder, add these examples, and cross reference it. So, yeah. uh a

lot of people kind of I'm not sure if I can use the word overengineer but they overthink they overthink what is AI but it's like as easy as just talking to it asking it

questions like how did you do this how did you do that and guiding it uh well of course if you have a good idea of what you want to achieve but it's very important to be able to break the

process into kind of uh smaller steps into building blocks so to say okay Next, after outline, what's the step?

So, now what we do is we look at the outline we've created and we ask Claude to find specific opportunities to mention relevant HF's products. Um, I

tried, you know, having this integrated into other steps and it was a bit hit and miss. And this is obviously

and miss. And this is obviously something that really matters to us because this is why we write content. We

want to talk about the product in context where it makes sense to do that.

So, this is a discrete stage. This will

do this every single time. Um

uh you know it's very simple cuz within the skill I actually have a kind of master list of HFS products and features which I asked Claude to create for me and then I updated and tweaked myself to

include like newer ones add some features. So it goes to that and it

features. So it goes to that and it looks at the outline. It says which of these can I contextually mention in this outline and have it make sense, have it be useful for the reader. And it just adds a little signpost for the next

step. Uh so that when it comes to

step. Uh so that when it comes to drafting, it knows to actually incorporate HFS into it.

You know, keyword explorer, that kind of thing.

And again, probably this is not something that people need to write start to finish themselves. Just drop

links to your landing pages to your video overviews. Ask it to analyze it

video overviews. Ask it to analyze it and tell it tell you what the product is, what is it, what is it good for, what are the top use cases, what are the like uh use cases for I don't know for

this area, for that area and then you just correct it. So yeah, it's actually those things are easier to create than than uh people might think.

Yeah, exactly. We've got site audit, rank tracker, content explorer. Uh

Claude did most of the heavy lifting here. I just reviewed it. Um and I added

here. I just reviewed it. Um and I added in some I need to add in like fire hose and things like that actually. Um but

again, Claude can do all this for you.

It's a fantastic diligent worker. Uh and

then after that is the drafting stage.

Now I think when most people would do like an AI content process, this is probably the only stage they would create. And certainly when I've talked

create. And certainly when I've talked to people, this is all they do. they

focus on what are the best prompts for making an article. But yeah, from all of our trial and error, I think having tons of steps for research and structure before you get to writing is what ends

up giving you the best outcome. Um,

and this is again similar to the writing rules we had in our previous GPT. It

just has some this is adapted from our own internal writer like style guide for writing. You know, use the problem,

writing. You know, use the problem, agitate, solution uh formula. Here's an

example of it in action as part of the introduction that works pretty well.

Some structural stuff that inverted pyramid uh always explain what and why uh all these very simple things and draft very well

draft is not a final step right is not the final step.

So what goes after draft?

So uh as we have a kind of verify claims stage um internal linking is very important for us and for SEO and also

making sure we have included useful up to-ate sources for everything that we do. So there is a particular step in

do. So there is a particular step in here that it actually goes through the draft and it looks for the claims things you know claims that the article is making that we would need to go out

and validate and it makes sure that it has an upto-date source for that or it update it reviews it to see if it's accurate or not. Um and actually I've

been working on this updating this skill because this is a big part of our content updating workflow. We want to go back to old articles, find all the claims, make sure they have the most up-to-date uh validation and accurate

stats for it. Uh so that's the next step of that process there.

And there's there's more steps after this.

Yeah, not too many more. Um so we have uh a preview stage. So at this point, I wanted to be able to look at the draft uh and sadly check it and see if I was happy with it. And it's not always I

don't like looking at markdown files like this. So, it actually generates a

like this. So, it actually generates a HTML file that is styled to look like the HF's blog. And I can then open that up in my browser just to like see what it would look like and feel like on the

blog so I can quickly review it from that point of view.

And the thing that still takes a ton of my time that I'm trying to work on is uh screenshots.

So much of our content is productled. It

involves using the HF's product.

Screenshots are so important for that.

At the moment, what this does is it will uh suggest a report uh that we can actually go and visit and take a screenshot of. And we actually have

screenshot of. And we actually have another skill that other people on the in the company have built which allows the claw to structure correct URLs for our reports. So it can actually generate

our reports. So it can actually generate a genuine report URL for you to visit in HFS and then I can take a screenshot of that.

So that's quite useful. I'm trying to automate that with some headless browser stuff and some screenshotting and that kind of thing. Um, but at the moment I spend as much time doing the screenshots

as I do actually editing, reviewing, generating. So that's a big part of it.

generating. So that's a big part of it.

Okay. Uh, since my my job on this podcast and in our calls is to essentially criticize everything you do.

What a fun job. people would would uh think I'm a terrible person, but it is what it is. To be honest, one one step I

I uh expected to see in this process is when you would uh kind of dictate to this system some of your thoughts of

where to take this article in free form and I would explain uh why ah you have it or something. You you're pointing something out.

I do indeed. Yeah, I kind of glossed over it. Um, I totally agree. Sometimes

over it. Um, I totally agree. Sometimes

you just want to provide a few sentences of thought or direction. You want to mention a specific product and you don't trust that it will do it itself. So, one

of the things I added recently was this context trigger. Um, so this right at

context trigger. Um, so this right at the get- go when you trigger the workflow, you can provide it with as many sentences of context as you would like and that is then used to shape and

inform the rest of the process. Um, so

often I'll say cover this topic or this topic or review this existing article and bring elements of that into it or mention this new product and that kind of thing and it's just a little directional nudge and again that seems

to be very useful for getting a good outcome from it. Uh, I think it's like a critical step in my opinion. Again uh we are still in the very early days of all

that. We're still experimenting and I

that. We're still experimenting and I like I have thought so many thoughts uh in regards to all this. So first of all, I think it's important to point out that what you just showed is a work in

progress because any any kind of skill, any kind of workflow that you build for yourself uh in cloud code or any other AI, it shouldn't be set in stone. Every

time you run it and every time you analyze the output whether total output or whether output of the steps and you don't like something you need to go and refine and you keep refining and

refining and you're basically teaching your AI workflow AI agent AI skills skill to do a better job and with every run it would get better and better. Uh

so this is the first point. The second

point I feel this uh this this step of giving it context is super important because it is what will essentially make

your content unique because again uh the reason why I was also surprised that you would let it run for 8 minutes and just generate something for you is because I

would expect that uh you would get a TLDDR file from the top competitors. you

would go through it and you would just like in free form uh I'm I'm using whisper flow this thing to dictate into into anywhere basically in text all the time and I would just click a button and

I would say oh like so I disagree with this part I think this part is good don't even mention this part is not important here is where I think you can and you can give it a lot of

instructions it's almost as when we had those uh content mastermind calls where we would discuss ideas and we would brainstorm where to take every idea In the same way that we were giving each

other feedback and uh kind of figuring out what angle is best to take uh with uh any given content idea in the same

way you can provide uh feedback to or context to AI and I feel it would it it typically would do a great job at doing this

and that very another very good point maybe I'll talk briefly about how I think conceptually this process should be use for content marketing generally like this is not the HF's content process going forward. It is not as

though everything we create has to come through this or will come through this.

Um we spend a lot of time writing stuff that is AI is still not very good at helping with things that require tons of thought and experience and unique perspectives and ideas that maybe other

people haven't even shared before. I

think this is really useful because, you know, we've written literally thousands of articles over the years. And what I see being really important for us going forward is having this well-maintained

library of evergreen search content. I

want to make sure we cover all the core topics that relate to our product and how to use it, keep them updated. And a

lot of times that is very simple, quite repetitive stuff like how many ways are there to do keyword research? Quite a

few as it turns out. So I think this is really good for topics you know we have tons of information documenting keyword research and all these kinds of topics that can be used to inform this process.

This is almost like you know doing our housekeeping for us in some sense. Um

it's not something that requires a ton of direct involvement guidance because we've already done that. We've written

dozens of articles on these topics that is using used to shape these articles now. Um, and you know, I've generated

now. Um, and you know, I've generated tons of articles from this that were I could have published and would have been fine, but I didn't know enough about them. I didn't think they were

them. I didn't think they were interesting enough, and I've chosen not to do that because I still deeply care about everything we publish and I'm I want to make sure we put out the best thing we can.

So yeah, I feel this process and the reason why kind of you let it run on itself with little output, it feels that it's best used to take some kind of what

we call a general knowledge topic and adapt it to us because one of the steps it pulls from our existing content and it finds what kind of unique stuff we

said. Then it finds the the way to uh

said. Then it finds the the way to uh include HFS in our use cases in this post. So basically for example there's

post. So basically for example there's plenty of information about link building but it doesn't necessarily share what we have shared about link building and it doesn't necessarily

makes good use of HF's tools when it comes to link building. So with this automated process this is where you don't need to write something from scratch. You can analyze existing

scratch. You can analyze existing content and AI can find a lot of information from our existing articles and from our tools to include in the post and yeah you have uh the post

ready. Am I right?

ready. Am I right?

Yeah, exactly that. Yeah. I like to think, is this a boring topic that I don't want to write? Um because we've covered it a thousand times. If so,

maybe it's a good candidate for the AI process, which is not everything we publish in that regard. Uh oh, you have you have some something else to say.

Yeah. So, very briefly, I because it kind of ties on to this. I'm also we built a content updating pipeline. This

is a bit newer. I'm still tinkering with this but in a similar you know we have yeah thousand published articles for example and it's very hard for human people to keep on top of that keep them

updated so we're working on a similar process here that is designed to basically periodically give you updated content to review and edit and approve and potentially publish um and very

similar thing there are basically three things this does it looks for claims that might be outdated so there's an old stat or something that doesn't make sense, Claude will review it and try and

find a new version of that and allow you to accept it if you want to. Um, you can find opportunities to add new hrefs product features. So, obviously some of

product features. So, obviously some of our articles were published like 8 years ago. They don't mention our latest

ago. They don't mention our latest products like fire hose or you know uh AI content helper. This can make recommendations for you. And lastly,

updating topic gaps. So this is where it looks at the SER and it says, "Is there anything that has other articles talk about that we don't? Perhaps we should draft a section for you to review and

edit and include." And it just makes, you know, very boring um unstructured process a bit more organized and a bit more um fun for people to engage with. I

think I I really really like where all of this is going because I think this is actually the future of how content is going to be created. And uh I wanted to

wrap wrap this up from a different perspective because you essentially shared a workflow of how uh to create content on what you call like a boring topic, something that has been covered

over and over and we just have like some unique spin or we want to cover this topic and include our products and services. I wanted to to share a quick

services. I wanted to to share a quick story uh from the other side when you want to create something completely unique. uh and that is uh so I'm in the

unique. uh and that is uh so I'm in the process of writing a book as I mentioned many times on this podcast already and uh just 8 months ago I was complaining

to uh a bunch of our team members that it is very hard for me to context switch because when I stop working on the book and I do some like projects uh inside

HFS and then I need to return to the book like a few weeks later I barely remember what I was writing about I barely remember my train of thought and it's almost like I need to upload all

the information from scratch and uh I think it was further who said uh why don't you just upload like all your chapters to AI and kind of ask it to guide you like AI would ask like a

journalist or a ghost writer who would be interviewing you asking you questions and would be kind of writing the book for you it was 8 months ago about 8 months ago and I said I cannot see how I

would be able to do that so back in the day we didn't have cloud code back in the day like Chad GPT just released their custom GPTs or something. I

couldn't see how I would upload like my entire book and be able to work with it.

Fast forward eight months and the last chapter of my book, I just finished the the draft. The last chapter I wrote it

the draft. The last chapter I wrote it with AI by dictating my ideas into cloud code and my process was I told it, okay, the name of the chapter is this. What is

going to happen is you're going to create a folder with my random dictations because I have a list of notes. what I want to say uh within this

notes. what I want to say uh within this chapter and those notes exist in the form of three words or one sentence basically talk about this or expand on

this idea and I would hit a button and I would just ramble. So there's this idea and they wanted to say blah blah blah and we like did this thing at HS and we

have this interesting story blah blah blah dictation over next idea and I was just uh rambling on each of my ideas. I

had a few dozen of them. Okay, it saved that to the folder. And I said, okay, I'm also like one talking when talking about those ideas, I was referencing a few things. Some of the things that I

few things. Some of the things that I discussed with some other marketing leaders on the podcast, some of the things that we actually covered on HF's blog, for example, we have an article about taste and I just said, "Oh, like

I'm talking about taste in in my chapter and you have my voice dictation with my ramblings about it, but we also wrote a nice post. Please include it as sources

nice post. Please include it as sources when talking about taste." So I gave AI I gave it all my dictations and I gave it all the resources that I remembered like different uh YouTube videos,

interviews, different articles that I want to reference etc. Even some LinkedIn posts that I saw from people who are sharing these ideas and then I

said okay now the general idea of this chapter is this. I'm trying to make a point that blah blah blah blah blah now you know like all my dictations now you

know all my resources all the stories I want to tell me how would you connect the dots how would you structure it so essentially create me an outline and it

would write me oh so I suggest that you lead with this story then it transitions well this and then this argument and then these things blah blah blah at which point I would say like I would

give it some feedback where change it or not or I would say sounds good to me write it and it would write a chapter for me and then I also like uh uploaded

to cloud code. I uh I downloaded from Google documents all my previous chapters and I said okay for each chapter create kind of a synopsis file what this chapter is about what are the

key arguments that I'm making and what is the TLDDR outline of a chapter what are the main stories and key ideas and I'm sharing so for each chapter it created this file kind of with a recap

of the chapter and then I said now refer to all the files of all the chapters and create me a synopsis of the book I want to know like what the book is about how it is structured and what is illogical

ical and it is so good. It's like it's literally like you're you're offloading some of your brain work to someone else like you have an external brain that

processes information for you. So this

is why kind of when when we started talking and when you shared that you created a system for uh creating uh blog post fast and you said that your productivity increased that you

published like three articles in a few days or something like that. I am

actually expecting that all of the content that we're going to create, it would go through AI that we will no longer manually write stuff. We would

just hit a button. We would ramble to AI what we want to say. We would point it at like whatever resources we want to use to make a point and it would help us

write even a better article because its ability to connect the dots and understand what you're saying is actually quite crazy. I'm very surprised

how well it was able to distill my ramblings into coherent ideas and connect the dots between them and organize it in a way where I'm like,

"Wo, this actually looks quite good." So

yeah, uh the process that that we just covered uh in in this podcast is uh mostly for kind of semi-automated content. you still want to like overlook

content. you still want to like overlook it and like like you said you have a step to give it context of where you want to take it and what's the unique angle and stuff like this but it's still like you're you're offloading the

majority of the work while I think going forward creating content yeah AI would act like a journalist an editor a ghost

writer and you would act as a source of ideas and opinions and people who don't have a good writing skill but have strong opinions would be able to publish their content fast. So, what are your

thoughts on this?

Yeah, I totally agree with that. Um, I

always find some people think that human creativity is too unique and magical and special that like AI could never help with it and never be a useful aid in that process. But actually, there's a

that process. But actually, there's a lot of mental drudgery we do when we're writing a book or an essay or anything like that. I think the ideas, the

like that. I think the ideas, the motivations, the experiences, the things we care about, that is still uniquely you in your book. still your book and your ideas. Yeah.

your ideas. Yeah.

But all you just sitting down for hours and shuffling these ideas about and working out what are the common themes that is something that AI is fantastic at doing.

Um yeah, if it can make these writing and creative processes more fun for us, then like that shouldn't be scary. I

think that should be fun. We'll be more prolific. We'll share more stuff.

prolific. We'll share more stuff.

There'll be more of our unique thoughts and ideas out in the world. Um, so if yeah, for all the kind of like sad drudgery and you know, are we automating careers and jobs away? Actually, we

could create more cool stuff than has ever existed before in human history.

It's totally possible. Now, I I like I like the word drudgery. I think what what AI does, it it literally eliminates drudgery because like I said, for me, it

was a pain to go back and to because I would need to read my entire chapter again to remember what I was saying there. And now I can say remind me what

there. And now I can say remind me what was the synopsis of the chapter where we left off. Uh which ideas need work. It

left off. Uh which ideas need work. It

would tell me all that and I'm like immediately I can continue working and I can pick up where we left off. So yeah,

let's let's not make it longer than uh than we need. Uh thanks a lot for sharing your uh process. Thanks a lot for as always letting me to jump in with

my thoughts and ideas. Uh I generally think we're on the right track with with uh these kinds of things and this is the future. Definitely this is the future of

future. Definitely this is the future of content marketing and content creation.

Thank you Ryan.

Thanks Tim.