Introduction

So hi everyone, my name is Camille and a quick intro. I'm nobody you've googled, no Wikipedia page, no viral tweet, no TED talk, but I might be someone you will remember in the next 15 minutes.

and you know those people on LinkedIn like X meta X Google X something I'm not that X and I felt like hey I want to be X basically because it sounds awesome

From Sales Burnout to Building AI Systems

and I felt like what I'm X actually so I'm X a soul -crushing spreadsheet in those slack saturated pipeline pyramids here I'm cursed KPI killing spectacular existential breakdown in sales and I felt like mmm if I'm going to speak to

people that don't want to hear from me and analyze basically dashboards that are a lie, I can do the same for a computer, at least it didn't hang up on me.

So I moved to engineering, turns out it's quite a thing you can have for building AI

systems because I don't build AI for fun, I build it because I felt every kind of broken process from the inside I know what to automate because I've done it all by hand pretty pretty badly so and one of the most boring things repetitive things

The LinkedIn Engagement Problem (and Why It Matters)

you do in go to market is LinkedIn engagement here's something most people don't know before buffer analyze 70 ,000 posts in thousand posts on LinkedIn and replying to comments on your own post boosts your engagement by 30 % and if you

comment on other people's posts this is basically engaged by the algorithm 5x meaning that it weights more than likes meaningful comments of course not great posts because those actually get like the prioritized and so that's problem I

Solution: Two Lightweight Automation Bots

solved with the two bots so on the left there is comments reply bot built for my clients scrapes all the comments under the latest post proposed replies based on their like voice and past feedback since an email every morning review tweak pulse five minutes instead of 45 minutes for a workflow and the second on

the right is engagement bot so pretty pretty similar idea for outbound but scrape post from the key people in your industry and scores by relevance generates common drafts, daily digests, like review, tweak, and post.

1Both run on GitHub Actions and are pretty free because their average cost monthly is around $10 and most costs are just scrapers and LLM usage which is very very cheap as you know.

How the Bots Fit Into a Full GTM Campaign (Results)

These bots are part of like full go -to -market campaign and on the very bottom of the of the slide you can see that I run for my clients and those campaigns like over 27 qualified

opportunities in six months each worth at least 6 ,000 pounds so in general like more than 150 thousand pounds in pipeline by this workflows but what you can get out of this presentation actually I know what you're thinking so

Making AI Writing Actually Good: Human-in-the-Loop Over Prompt Bloat

a writing LinkedIn comments that sounds weak like there are tools for that already you can spot them from a mile away and they're terrible so ask myself how to do make this actually work because the obvious approach doesn't

yeah so that edit arm you can sit down and write a prompt that that covers every edge every kind of sound every every keyword your client would like to communicate and stuff you could dump all their previous posts as context but this

This is just noise.

The Feedback Loop: Propose → Review → Save Edits → Improve

1What actually works is active human in the loop. AI proposes human reviews, edits posts. This edit gets saved.

Do this 10, 20, 50 times, and the model analyzes the feedback, refine the prompt, and it keeps getting better and better, and it's a never -ending process.

So I did a deep research on this, how to do HETL, so human in the loop, workflows with AI. and I want to proceed present you the key insights from from my research

Operational Rules for HETL Workflows

basically so the the one that 25 to 5 rule 25 % of your workflow should be run without the human but the 5 % of human input that's the 80 % like of the of the value the value that the human should actually add to set your confidence

threshold not to 70 % to sorry to 70 % not 95 % if you set it at 95 % the AI routes most cases back to you and automation that doesn't automate actually so 70 plus spot checking every 10th basically output beats perfectionism every time

free every approval gets gate needs a timeout otherwise someone forget to click and basically your workflow dies in silence so no decision 24 hours escalate for don't out make a broken process that's super important a

amplifies what's already there so if the process is broken AI makes it fall faster basically and and five sometimes the human is the problem actually so a

fortune study found AI alone outperform AI plus human in low and advertising people second -guess correct a decision because they because we give a bad report to AI that it's hallucinates and stuff so that's very interesting six the

framework for every step you think about to automate you should ask yourself three questions what's the cost of a mistake how repeatable it is and does a human actually add value here if the answers are low very and no automate

made without the guilty trip and probably so you want to build it on

A Practical, Mostly-Free Toolkit to Build This Yourself

GitHub Actions as the Always-On Automation Runner

yourself like this here are your free toolkit as the meme says free real estate so the tools I've used one github actions like 2003 credits minutes per month to launch your Linux machine or MacBook machine Mac OS machine that's as your server, your brand, your pipeline, your scrapers,

your daily automations, free.

Claude Skills and Session Reflection to Improve Instructions

Two, if you're a Cloud Code fan, you probably know about the latest updates, so skills and insights for Cloud Code.

Skills are just repeatable tasks you can automate with Cloud Code insights.

Basically, what Cloud did is that they upload a skill which is a retrospective. So Cloud Code analyzes how you use it and makes you insights and suggestions how to use it better and so watches

itself but I feel like hey why I have to spawn a comment for that it shouldn't do it on automate so I found two projects that do exactly this cloud deception the

first one there is a QR code to the github repo your cloud or during a session it captures those corrections and updates you in your instructions and and so the same mistake doesn't happen again and also cloud reflect your correct cloud and also It captures this.

I'm sorry cloud deception

basically checks if you discovered something non -obvious or work around the pattern at the beginning and creates a skills out of it and the second one basically sees how you correct cloud and makes it a rule basically, so it never happens again and

And the other things worth to use,

Data and Storage: Supabase, Apify, and RapidAPI

Superbase, super free for the free tier. You have a lot of storage there, is Postgres database.

And also if you want any data, Appify and Rapid API, they're really, really good at it. You will find whatever data you want for those scrapers.

Yeah.

Build vs. Buy: Why Custom Bots Beat Off-the-Shelf Tools (Sometimes)

So, and also looking at the agents I built, you can tell, hey, why did you build that actually? because there are services that you can pay that does the same like buffer and

the first thing most of those services don't allows you to play with your LLM you cannot change the prompt you don't know you cannot put you know the temperature and stuff I don't like that it's like you feel like a black box client in those services so I always want to see what I put what's the prompt

what I prompted to do and what's the next step and the second thing and go to to market engineering is like the things I built today maybe I won't use it tomorrow or the next week so we have to eat it very quickly and this cloud code

plus sub super base plus github actions allows me to build something that is really similar to buffer maybe has like 10 % of features but allows me to quickly test if this is something for me and also knows what I like what I don't like

so that I don't have to pay I know buffer is from $20 to 100 for for a subscription it's not a lot but I build that in two hours and I've checked if it works if it works for my clients what ROI I can get for it and then I can decide what tier of the more upgraded more more top tier the kind of tool I

need yeah so actually actually that's it for me so I left a lot of time so yeah just remember those rules for H ETL and happy to if you want to talk about it it. Happy to connect to LinkedIn and also if you have any questions, yeah, let's go.

Q&A and Wrap-Up

Yacek? So is Cloud Subtion, is that right in Cloud? Is that a plug -in or is that... It's a skill. It's a skill.

So whenever you have any long session, it automatically asks, hey, can I run a reflection on your session to just correct Cloud MD, Cloud MDs. If you don't use code, it's like a main brain thing, main markdown file that basically tells how it should act. And every session I do, he asks me, hey, can I run a reflect to see what mistakes I did and how you corrected me? So it never happens again. Yeah, it works like that.

Staying Safe with LinkedIn Automation

I have a general question about the linking automations. Yeah. So how to not be called by the linking as a bot? Sure. Do you learn something special for you or do you just use tools to do that?

So I use tools that does it and that have regulations and know how to do it properly and that we won't get banned, basically. But, yeah, that's the best. Basically, so find a tool that knows how to do it because I know connecting with LinkedIn API is really troublesome and they don't want, like, standard users to do it. You need to have a business and stuff and you have to work with them, work on the regulations and stuff. So there are already very cheap services for that, yeah.

Okay, so I think that's it. One request from you. Can I take a picture with you and post it on LinkedIn? So from the stage, is it okay? Okay, so let's do the photo, selfie, yeah, the inception of selfie. Can you look happy that I did a good job? Okay, yeah, thank you so much. Yeah, we got it. Okay, thank you so much.

Conclusion

if somebody wants a presentation let me know a link and I will send it thank you so much and I wish you a pleasant evening