Support Stack E12: Rewriting 235 Articles for Fin
Support Stack – Episode 12 with Gabriela Passeri
What actually happens when you switch on Fin with a legacy help centre?
For most teams, the answer is frustration.
In this episode of Support Stack, I’m joined by Gabriela Passeri, Knowledge Manager at AutoDS, to break down how her team rebuilt their knowledge base from the ground up to make it truly AI-ready.
They didn’t copy and paste 235 existing articles into Intercom.
They reduced them to 50.
They merged fragmented topics into complete, end-to-end guides.
They rewrote content to remove vague language.
They standardised product terminology.
They documented every error message clearly.
They built guardrails inside Claude to enforce consistency.
And the result? Extremely high early accuracy and a strong foundation for Fin to perform well.
What we explore
Why content written “for humans” often fails AI
How to structure articles so Fin can interpret them properly
Writing for three audiences at once: users, agents, and AI
Why specificity matters more than ever
When to fix content vs when to lean on Guidance or workflows
How beginners in your user base change your writing strategy
How Claude Projects can enforce consistency across a growing knowledge team
One of the most important takeaways: content is not just for customers anymore.
Your knowledge base now powers:
Your users
Your support team
Your AI agent
If the foundation is messy, Fin will struggle.
If the foundation is structured, precise, and intentional, everything gets easier.
This episode is a practical look at what that rebuild actually looks like.
If you’re leading a support team using Intercom and Fin, I share daily, practical insights on AI-ready support, resolution rate strategy, and evolving support teams.
Here’s Gabriela’s excellent, ready-to-use prompts: https://docs.google.com/document/d/1Q7eFGcrwHi70SyOltGKw6emiIThh82au9hQqFGLSOxo/edit?tab=t.0#heading=h.u9e4hny14af
Subscribe to my daily email here to get notified of future episodes.
Episode Transcript
Conor Pendergrast (00:00)
Hello and welcome to episode 15 of Support Stack.
Well, Gabriela Passeri, welcome to Support Stack. How are you doing today?
Gabriela Passeri (00:14)
Thank you Connor, I'm doing very well. I hope you do good too.
Conor Pendergrast (00:18)
I am doing extremely well, thank you very much, and all the better for having you as our guest. So you are knowledge manager at AutoDS and you, I said this last episode, if you have missed last episode, then dear watcher, you're not subscribed. So you've got to click that subscribe button, it's right at the top. And Gabriela was on with us last time talking about how she added a glossary, added a series of glossaries for FinAI agent to understand.
customer language better and also understand the product better as well as screenshots and images. So it's a fascinating episode. But today we're talking about knowledge in general. And that's obviously your shtick and your whole job, is great, but give us some background here. what it like you made the switch, your AutoDS made the switch to Intercom and Intercom's Fin.ai agent.
But you had existing articles. Did you just copy paste them into intercom and intercoms help center? Like every single one just went straight in there. What was your approach there?
Gabriela Passeri (01:17)
Of course not, Conor Because we had articles that had no AI written. No, no, no, no. Everything was a mess for AI. So we decided to change the way we are writing the articles for Fin to understand. Not just because the article is not good, but because they don't have a way for Fin to read exactly the way that we need it.
Conor Pendergrast (01:43)
Yeah. Yes. So I think, I think it's an important thing to know it is that like, if you are switching to intercom, or even if you're just switching from not using Finn to using Finn within intercom, you have an opportunity to work out the best way to have your knowledge center set up. And also the best way to write your content. Cause there are differences in how to write your content for AI, then how to write your content for humans. could be a lot sloppier when you're writing for just for humans.
Because especially when you have a well-trained teammate, they will just understand a lot easier and be a lot more critical and a lot less literal with the writing. So you went through a process of trimming down and overhauling your help articles, is that correct?
Gabriela Passeri (02:12)
Yeah, it's a shame.
Yes, it is correct. And also because we understand that the way that we are writing the articles was for users, but not quite good because not covering everything that the user should know in the same article. So the first step was to merge the articles into one with the same topic related.
Conor Pendergrast (02:53)
Mm-hmm.
Gabriela Passeri (02:53)
And
this way a user could understand in the full page of everything that they should know about it.
Conor Pendergrast (03:01)
Interesting. OK, so beforehand you had a I understand you had a large number of articles. So you had 235 or so articles beforehand. And then what you did was 235. And so you stripped that right down and merged a lot of content into longer topics. Now you have 55. Great. And so how?
Gabriela Passeri (03:19)
Yeah, we now have 50.
Conor Pendergrast (03:28)
Like what approach did you take? Was it just saying like, I know, were you already really, really aware of all the content in there? I imagine that as the knowledge manager, you probably knew everything about it and you just knew, okay, these 20 topics, these 20 articles, they can go into this one or maybe two different pages. that how you were doing it? Were you using a more structured approach? What was the approach? Yeah.
Gabriela Passeri (03:52)
It ⁓ is because of
my background. I was an agent. Like before I became a knowledge manager, I was an agent and I saw the users need help in something. I had the knowledge and also have the perspective of the agent. This way I could realize what we needed beforehand.
Conor Pendergrast (03:57)
Yes. Yes.
Mm-hmm.
Yeah, super. you had, I think that's a really important point is that knowledge managers can be well equipped with a lot of experience based on actually interacting with customers. And you get that muscle memory for the articles that you refer to a lot and the articles then, okay, well I'm sending that article. I should send this article as well. So you're sending like ⁓ a bunch of documents and over time, you know, okay, well customers got confused about this because I didn't send this article. So now I need to send that one more often.
And so that's really useful to remind people of is when you're, if you don't have a dedicate, go.
Gabriela Passeri (04:45)
Yeah, I agree. ⁓
I really listened to the user because I was an agent. So I realized that the article was not good enough to cover everything that they should know. So this was the first time. I think it was the, like the click, you know, I need.
Conor Pendergrast (04:56)
Mm-hmm.
Yeah. Yep.
Gabriela Passeri (05:04)
to change this, I need to do something about this because it's not good enough for them to understand and also not good enough for the Asians to understand because our platform is very complex. So we need to really deep dive in everything that we are talking about.
Conor Pendergrast (05:19)
Yeah. So that's, I think that's an important point to dwell on for a second as well, which is when you're writing a help article or when you're creating a collection of articles or a whole knowledge base, you're not just writing for one audience. You're writing for your customers. You're writing for your AI agent, you're writing for Finn, and you're writing for your teammates as well. So all your customer support agents. And even within then, within that group, those groupings, like you probably have different persona within your customer base.
You know, maybe you have not so much with AutoDS probably, but maybe you have some teams where there are different roles or different, you know, there's end users and there's administrators and maybe there's like finance admins and all these different things. So you've got, you've got to keep all of those different roles into account. But I think is it fair to say that like one of the main drivers behind your project was to deliver a better support knowledge base for Fin.
Gabriela Passeri (06:16)
not just for Femme, also for the users because what it is obvious to us as an agent, as the internal team from the company is not obvious to the user. So you need to write everything down, like the step by step, the little details, you know? So we were thinking about the user experience ⁓ with the article as well, not just Femme to read the article.
Conor Pendergrast (06:17)
Not just for fun, yeah.
Yeah, okay. So let's get a little, I like in support stack, the whole point of this series, I think, and these videos is to show and not just tell. So what can you show us here, Gabriela? What approach did you take when you were writing these articles? What can we look at?
Gabriela Passeri (06:56)
So I can show a little bit
how to use Claude. But also, I want to show
Conor Pendergrast (07:01)
Lovely. So this is Claude.
This is the Claude desktop application, is it?
Gabriela Passeri (07:05)
Yes, this is the desktop.
Conor Pendergrast (07:07)
Perfect.
Lovely. And so we are looking at projects at the moment. that's exciting. Can you, so if someone doesn't know what a project is, how would you describe a project?
Gabriela Passeri (07:17)
It is like a folder that everything that you are instructing here will follow the same for each chat you create inside the project. That's why we use projects because we want, for example, to write on Help Center the same language, the same way of to do, not to do, everything. So if you click on the Help Center,
Conor Pendergrast (07:22)
Hmm.
It's super.
Gabriela Passeri (07:38)
So here we create the instructions for Claude to create an article. So, or create or rewrite because depends on what you need.
Conor Pendergrast (07:38)
Cool.
Gabriela Passeri (07:49)
If you are going to rewrite, you need to give the old link for Claude to have access. So everything that you want Claude to have access, you need to add into these instructions. So you need to add what you want. you're one, hitting two, everything you have.
Conor Pendergrast (07:54)
Of course.
Gabriela Passeri (08:06)
fetch all and read all the content for
URL. if you're, for example, giving Claude another article, put the URL. If you're giving the guidelines you are using for Fin because it is good for Claude to know what are the guidance Fin are using. So it is good to also attach here.
Conor Pendergrast (08:16)
Mm-hmm.
Mm-hmm.
Gabriela Passeri (08:30)
We are giving more and more and more about do not do this, do not use this, do not... So everything we are guarding. ⁓ One thing that is important for the AI for Fin, it is to guide me on what to do and what to not do. So we are guiding Claude first, okay? So...
Conor Pendergrast (08:43)
Mm-hmm.
Gabriela Passeri (08:51)
Do not use vague reference like this, that, the system, because this is not clear enough for Fin to understand what we are referring to. So every time that you build like the paragraph, say the name of the feature. So it's not this feature. You need to talk about, oh, auto DS or order credits is this. So be specific what you are talking about, not just you.
is that an it
Conor Pendergrast (09:22)
Yeah, so you're using proper terms and proper words and proper nouns in all those cases, not just this feature. That's great. Okay. This is
Gabriela Passeri (09:29)
And Fink does
not read like sessions, hitting one, hitting two. Fink cannot read. So every time that you create a session for the article, you need to add a description of what is that session first.
Conor Pendergrast (09:37)
Mm-hmm.
Yep,
you repeat the heading and you say like, okay, here's a new section. This section is about, I don't know, priority shipping or something along those lines. Here's how you understand, here's how you explain how to get priority shipping or how to set up your shipping options, for example.
Gabriela Passeri (09:49)
You
And this is for everything that you're doing. For example, for errors inside your own platform, you must have some error message. You need to write every error, the same text thing you're adding to your platform. You need to add the article and what to do and why you are experiencing this ⁓ error message because users wants to know why. So first, why, and then what to do.
Conor Pendergrast (10:06)
Mm-hmm.
Yeah, lots of error messages, yep.
Mm-hmm.
Yep. Excellent. Okay. I'm curious. So, ⁓ last, last episode, we looked a bit at the glossary side of things. ⁓ and so one thing is, one of the things you said was this, the publicly facing articles should always be in the platform language and the product language to make sure that it's clear and consistent that way. ⁓ so are you codifying this, that in here as well? But I, I guess it's already there because of the fact that you have your
your source content is all the old content, that correct?
Gabriela Passeri (11:01)
Yeah, we do have the old content, like in the beginning, when we were rewriting. But now we do not have it. We only have the published one with the 55 articles we have inside Intercom. So for user facing, we write everything the same terminology as the platform, because Fee needs to understand what they're talking about. And the glossary is using the user terminology.
Conor Pendergrast (11:05)
Mm-hmm.
Of course.
Gabriela Passeri (11:27)
So every time that you are building a project for a Claude, you need to add everything, ⁓ all the features you have. So we also attach the Glossary here. Every thing, explain each feature, you know? So Claude knows everything about AutoDS now.
Conor Pendergrast (11:38)
Super. Super.
Gabriela Passeri (11:45)
And every time that we need something, Claude will build using the instructions we added. Like for example, in the last episode we did, we talked about the managed balance. So we guide Claude to not use balance as a word, just use balance as a feature, the name of the feature only. Never use for anything else, because this can go build feed.
Conor Pendergrast (12:09)
Yes.
Yeah, yeah, that's fantastic. So this looks like a pretty long set of instructions. Do you have examples for Claude in there as well, so that Claude knows what an article should look like, a good article should look like? yeah.
Gabriela Passeri (12:25)
Yes,
we do have. We shared a good Excel workflow with him, like best practice. Do not do this, do not do that. But it is a long one. You can have a document that shares everything. Like we have templates for you. We have the guidelines to use Claude. It is a good one. Check the document.
Conor Pendergrast (12:29)
Mm-hmm.
It's really long, it's amazing.
Mm-hmm. Yep.
Yes. So if you, again, viewer, look down in the little description box, there's a link there to the Google Doc that Gabriela has very kindly put together, which covers both this, the instructions here, as well as the setup for glossary. And you can use both to create your own glossaries and create your own fantastically well-prepared articles. that's a, I think that's a real bonus and a really helpful thing. So.
Talking
the practicalities of this then, when you, so you've got these instructions set up. We won't show it here because it takes quite a while I imagine, but what you would do is you would come and say, hi, I would like you, hi Claude, I would like you to rewrite these articles into a new single help article. you, based on your amazing experience as a customer support agent and now your experience as a knowledge manager.
you would go through and you would pick out four or five links, for example, and place them in a chat and say, Hey, here's the next topic to go through. And then Claude would just go through the content of those pages using your detailed instructions and your examples, and then create the new content. Is that, that's correct?
Gabriela Passeri (13:58)
Yeah, we create, first we create a Confluence article, guiding ⁓ the format, we want. Was the first step, like we want the headman to be like this, do not use like double dashes, everything that we want inside this guideline. We create the guideline only for Claude. Then we use the instruction here and we start using. When we start using, we realize we need more.
Conor Pendergrast (14:20)
Mm-hmm.
Gabriela Passeri (14:26)
And every time that we realize that, we change the instructions again and change the guidelines to have this as a backup. So from that, we understand based on the conversations we check. So we need also keep this updated. Every time we are ⁓ giving new instructions to Claude, everything here, we need to be super updated.
Conor Pendergrast (14:30)
Yeah, of course.
Yes.
Yeah. And it's so, is this a project that's across multiple people? Is it, this is your one, your one project? This is like a team project.
Gabriela Passeri (15:01)
When I was building the department, because the department was new, I was the first one handling the department. So I create all the guidelines, like SOPs, the guidelines to write an article in a good way, because I was ⁓ graphic designer back in the past. So...
I like to do UX for writing as well. So I create the logical reason for each article based on my experience as an agent. And then from this document I created, we had now two technical writers. So I would build in, I have my girls are amazing. They're doing a very good job.
Conor Pendergrast (15:25)
Mm-hmm.
Yeah.
Gabriela Passeri (15:44)
Because I create this one and they are keeping this updated or Fin is not using this well. So I need to change the instruction for Claude. I need to add this. We need to have more people. I need to have more people, the two girls, they handle the project, like the migration to intercom. So they wrote everything that we need. We started with the user facing and now we are writing articles for copilot is agent facing.
Conor Pendergrast (15:53)
Fantastic.
Ha
Yeah, yeah, that's amazing. And so I think I've seen, I can't remember the numbers around it, but this, the work you've done here has set a really strong foundation for Fin, right? Like you, if I'm understanding just from content and the, this content and the glossaries alone, and without doing much apart from that, without like touching on data connectors that much, without touching on Fin procedures that much, you've got a pretty substantial resolution rate, right?
Gabriela Passeri (16:43)
Yeah, we had 97 % accuracy in the first week. It was amazing because the content was built and we are having more material of the procedures, our workflows and guidance. So we, and this is the good part, I changed the team now works closely to the team with the workflows, the procedures and guidance because we talk the same language. We need to understand
Conor Pendergrast (16:47)
Yeah, great. Yeah.
Mm-hmm.
Of course.
Gabriela Passeri (17:11)
Like for example, if it's content related or guidance related, because we can confuse sometimes if the guidance or content. And we try first for content. We always try first for content. And if it's not a helping thing, then we go to the guidance. The team for guidance workflows and procedures, they have a lot of work. It is a lot, not...
Conor Pendergrast (17:30)
Yep, that's fantastic.
Gabriela Passeri (17:39)
This is good for knowledge ⁓ base, but we need someone to take the part of the sex operation for sure.
Conor Pendergrast (17:48)
Yeah, it's, I mean, that's, that's my better bread and butter. Isn't it? It's like, that's, I spend a lot of my working hours in data connectors and, ⁓ and proceed tasks or procedures and simulations. And it is interesting. Like the great, a great knowledge foundation can get you really far. All of it. A lot of it depends on your customer type. A lot of it depends on your communications, communication approaches, what people prioritize, what kind of products you have, what kind of company you have.
but you can get like 55, 60 % resolution rates. I know, God, there was a previous episode ⁓ with Toby from Pupil Progress. I think they got 75 % resolution rate ⁓ based purely on content, guidance, and audiences, and it's fantastic, but I have also no other people who...
Gabriela Passeri (18:31)
This is it.
Conor Pendergrast (18:40)
struggled to hit 55 % with content and it's been a slog of data connectors and procedures to get the rest of the way to 64, 65%. So a lot of it does depend on the products you're building and like who's interacting with it and how things go wrong. Cause fundamentally that's the challenge is like most of the time people aren't reaching out just for fun. They're reaching out cause something's gone wrong. But this is like, this is such a strong foundation. It's fantastic.
Gabriela Passeri (19:07)
Yeah, you need to understand your user as well. Like for us, we do have beginners, like a lot of beginners. So to talk to them, you need to talk like a child. You need to guide them like a child. So every step in the way you need to add, like, do this, go there, do that, everything, little details.
because also our platform is very good, but sometimes does not provide the text. So it is an icon. It's not that intuitive, you know? You need to make it more intuitive for them as well.
Conor Pendergrast (19:42)
Yeah. Yep. Yep. And then that comes back to what you were saying in the last episode of the Glossary Contains screenshots and guides Finn to understand what's going on. That's first. Actually, I'm curious when you were creating the, when you were doing this work to create the help center and prepare that, did you have instructions in there about screenshots as well and like recommendations? Cause obviously Claude isn't going to be able to generate the screenshots. ⁓ but you probably have the ability for Claude to be like, ⁓ put a screenshot in here.
Gabriela Passeri (20:05)
No.
Conor Pendergrast (20:10)
Like here's a placeholder for screenshots. Yeah.
Gabriela Passeri (20:12)
Yeah. Like for example,
for main ones, we, our goal was not to put every single screenshot step by step because there will be like this and the experience to read the article will be bad. So we added to call that each session should have at least a main one explaining the page. So yes, we do have, ⁓ since Claude cannot take screenshots or do anything about the images.
Conor Pendergrast (20:21)
Yes.
Mm-hmm.
Mm-hmm. Yes.
Gabriela Passeri (20:39)
We did it ourselves with a guidelines for color, how to take a screenshot. You need to show the whole screen because the user will not find everything you are showing if you're not taking the big picture. So yeah, we do have.
Conor Pendergrast (20:52)
That's great.
So, so your focus was on the sections should have screenshots, not every single step. That makes sense. And I, I think also the challenge of screenshots and there are some products out there that try to solve for this, which is useful to know, but the challenge of screenshots is you have to maintain screenshots. The product will change and you have a slight UI change that makes
very confusing that you might not be aware of and then two months later Finn has sent 200 screenshots which are out of date. that's images are powerful and with great power comes great responsibility.
Gabriela Passeri (21:23)
We have to have a good.
Of course, we do have a good process for product releases. So everything that we are doing, we are aware and we know that what we need to update. So yes.
Conor Pendergrast (21:33)
Yeah. Good.
Yeah.
Good. I have said it before and I'll say it again, like especially, I mean, it was always the case, but I think finally people are understanding in now that we're getting AI agents in support organizations, support organizations do not exist in a silo. Your AI agent will not exist in a silo. So you have to have, yeah, you're kind of, you're in a lot of ways people redesign their whole companies to make great support experiences. And yeah, like that.
Gabriela Passeri (21:58)
No. I'm miracle.
Conor Pendergrast (22:08)
That kind of has to happen if marketing and sales and engineering and operations and every single team is not aligned, then it's gonna be a real, really big challenge to get an effective AI agent in the same way that it was a big challenge. And I'd say in most companies never really got resolved pre AI agents with outsourced support teams and in-house support teams.
Gabriela Passeri (22:31)
Of course you need to have people to teach the AI, you know, because it's not a miracle. It's not something that you put there and start working because AI it is independent, not independent because you need to create prompts for AI to understand your own content, your own workflows, categorization, everything. And AI will not work itself like not independent. So we need people and.
Conor Pendergrast (22:35)
Mm-hmm.
Yeah.
Gabriela Passeri (22:58)
The most, I think the most ⁓ issue they're trying to ⁓ resolve it is AI will take agents jobs. No, because AI will only be capable to ⁓ help with the common questions. The deep investigations from the agent side and this is our goal to have ⁓ more.
time to resolve big issues and not small things that AI can resolve for the user. And also because we will open more ⁓ positions like inside my team. We are a new department with new people. So we need more people to build AI and it is a balance, you know.
Conor Pendergrast (23:23)
Mm-hmm.
Absolutely. Yeah. The roles will shift. Yeah. I did a talk at Elevate CX about that last October. It was on Halloween actually, exactly about your sorts of roles, the roles that are evolving, the roles that are being created, the roles that are available for us all to work through as AI agents become more common in the workspace and in support for organizations. Well, Gabriela.
Gabriela Passeri (24:04)
We're busy.
Conor Pendergrast (24:06)
Yeah, exactly. This was great. as I said, dear viewer, click the subscribe button and also take a look at the link down below. You've got a link to Gabriela's guide on how to set this up and how to set up the glossary that we talking about last time.
Gabriela if people are fascinated by your work, where can they find you on the internet?
Gabriela Passeri (24:27)
Please go to the LinkedIn search for Gabriela Passeri and I will be happy to talk to you about it and we can discuss everything and help. I'm here.
Conor Pendergrast (24:30)
Lovely.
Perfect, thank you very much. Well, dear watcher, if you're still here, I hope you are. Click the like button, please feed my ravenous ego. And you can also find my work at customersuccess.cx. That's where I write my week daily email to customer support leaders using FinAI agent, all about how FinAI agent can be taught to behave incredibly well and also how our customer support teams and organizations are evolving over time. Yes, keep watching.
the videos I guess. Thank you very much. Goodbye. Bye.
Gabriela Passeri (25:05)
Thank you. Bye.