We Ran an AI Hackathon for Our Content Team. Here’s What We Built with Agent A
If you’ve been on LinkedIn lately, you’ve probably seen the AI-flex posts.
Some marketer automated their entire workflow. Cut their week to four hours and cloned their voice. Built an agent that drafts, ships, and reports on itself. Maybe whitened their teeth too.
Elena Verna, CMO at Lovable, called it out perfectly:
“Everyone has a system, a stack, a workflow that supposedly changed their life, cured burnout, and maybe whitened their teeth. It creates the illusion that everyone else has it figured out. So you hesitate to ask basic questions, because it feels like you’re the only one who doesn’t get it.”
Beyond LinkedIn, there’s a quieter pressure: every content team I know is being told from above to “use AI more”. So that the team can cut costs, ship faster, and be more productive. Not just 10X, but 100X.
The problem is “use AI more” isn’t a brief. It creates anxiety and not direction. So most marketers I know are stuck in this weird middle: they know AI could help, they don’t know where to start, and they don’t want to admit it on LinkedIn.
This is silly because content and SEO teams are sitting on a pile of obvious automation candidates. For example: research, updating posts, monitoring competitors, refreshing data, finding ideas, drafting briefs, and formatting for WordPress.
So instead of telling everyone on the Ahrefs content team to “use AI more,” we tried something more concrete.
We ran an AI hackathon with Agent A, our AI marketing agent.
How we ran the hackathon
The week before the hackathon, Ryan Law, our Director of Content Marketing, dropped a message in our team Slack: no writing this week. Instead, spend the entire week building your own AI content system to automate or speed up whatever part of your role you find most painful.
The “rules”, if you will:
On Monday, share what you’re trying to build.
During the week, build it in our shared Agent A workspace.
On Friday, share what you built, why you built it, and how it works.
Ryan also gave us one important constraint: The more specific your goal, the better the outcome.
The point was not to create perfect products in a week. It was to force everyone to pick a real bottleneck and build a useful v1.
Agent A gave us the place to do that. Especially since it’s connected to Ahrefs data where we could build around actual content and SEO workflows.
blog pipeline from Claude Code to Agent A without a hitch:
While Louise built her own Editorial pipeline: brief → outline → draft → edit → polish → verify → publish, with scrapbook context fed into every stage.
Each stage’s output is editable before moving on, and after it finishes there’s a Refine mode, a chat loop where Louise can ask for changes (“tighten the intro”, “swap this example”) and adopt or revert each one individually.
My Data Refresh automates the surprisingly painful quarterly chore of updating our data-driven posts (top Google searches, top Google questions, and so on). It pulls fresh data, filters it, and hands me TablePress-ready output.
My Press Release Generator turns a blog URL or product-feature note into a press release; goal is to plug it into our data-studies category so every new study auto-generates one.
Louise’s WP Processor takes a finished draft and returns WordPress-ready HTML with internal links and formatting handled.
None of these are sexy. All of them claw back hours.
The plumbing nobody notices
The thing that quietly impressed me most isn’t a tool.
It’s the pattern Mateusz wired through Scrapbook, Notes, and Source of Truth: every repo has an index.json that auto-updates whenever a file is created, edited, or deleted.
From that index, a lightweight reference file gets regenerated, a plain-text summary the agent reads at the start of any conversation. The agent knows what exists without fetching anything, and only pulls full content when it actually needs it.
What we learned from the week
A few things came out of the demos on Friday that we didn’t see coming on Monday.
Building with Agent A is addictive in a way using ChatGPT isn’t
As Mateusz said:
“This tool expands what feels possible, and it’s addictive. You keep thinking about what else you could build, even beyond SEO.”
This was how Mateusz ended up with tools like Scrapbook, his very own inspirations clipping tool. Paste any URL or raw text, and Agent A will read it and generate a structured note with a summary, key bullet points, specific claims, data points, and three article ideas inspired by the content.
It’s not directly SEO-related but it’s a base for him to draft his next thought leadership piece.
That’s what “use AI more” can’t capture. Using ChatGPT feels like asking a smart friend for a favour. Building a tool feels like hiring one. Once you’ve hired one and watched it work, you start looking around your week for the next thing to hand off.
The best tools wrapped around things people already did
None of the standout projects asked anyone to invent a new workflow from scratch.
We were already saving LinkedIn posts; SavedIn made the saves usable.
We were already collecting URLs; Scrapbook gave them structure.
We were already lurking on Reddit; the listener turned the lurking into a weekly report.
We were already refreshing data posts every quarter; Data Refresh just made the refresh take an hour instead of a day.
Don’t build a tool that requires a new habit. Build the one that makes an existing habit faster.
Memory and context matters more than word generation
The big unlock wasn’t “AI can write.” Everyone knows that.
It was that the agent could pull up the right facts, like past drafts, saved research, our internal style guide, what we already rank for, without us pasting them in every time.
Tools like Source of Truth, Scrapbook, SavedIn, Notes, the GitHub-backed indexes, Louise’s writing-sample library, the editorial-style skill, none of these generate content. They capture, organise, and retrieve context.
The drafts that come out of pipelines hooked into them are markedly better than drafts from pipelines that aren’t. If you’re picking one thing to copy from this hackathon, copy the memory layer first. The writing tools improve themselves once the memory exists.
Old builds port over fast
Louise had already prototyped pieces of her workflow on Lovable, and was bracing for a painful rebuild. She got the opposite:
“It’s very easy to move a project from another platform like Lovable and rebuild it in Agent A. Just export the code and Agent A instantly rebuilds it.”
So if you’ve already started building somewhere else, you don’t lose the work. You just plug it in next to Ahrefs data.
How to run your own AI content hackathon with Agent A
If your team is stuck in the “use AI more” fog, run a version of this. Here’s the playbook, in the order it actually has to happen.
1. Pick one team
Our hackathon was only four people. All on the content team. We didn’t invite anyone else from sales or product marketing to join in.
You’d want to resist the urge to make it cross-functional on round one. Twenty people across three departments turns the hackathon into a series of Zoom calls and meetings. That defeats the purpose of a hackathon, which is to build.
Pick the team with the most repeatable, painful workflows. Content, SEO, ops, support, lifecycle marketing — anywhere people do roughly the same thing every week. Roll it out wider after you have demos to point at.
2. Block the full week on calendars
This is the one that quietly kills most “innovation weeks.” Don’t ask people to build “alongside” their normal work. They’ll default to the normal work.
Ryan cleared our week the Friday before: no posts, no edits, no meetings outside the hackathon, OOO replies on Slack. If you genuinely can’t spare five days, do three. Don’t do one.
3. Have everyone write a frustrations list before they touch the agent
I’ll be honest: We didn’t do this for our hackathon. But I did this for myself personally and found it helpful.
Because the list of what you could build is infinite. Between that and “use AI more”, you can be caught in a panic and end up doing nothing. So, having a list of frustrations made tackling the hackathon easier.
So, you’d want to list down the things in your job that you keep doing manually that you wish you didn’t have to. That’s how I came up with my Data Refresh tool. It was because something that looked so simple on paper took me surprisingly long to do.
Two rules:
Be specific. Not “research”, but “I spend two hours every Monday going through my LinkedIn saves and pasting the good ones into a doc.”
Be honest. Boring chores count. The most-used tools we built came from chores, not from anyone’s clever AI idea.
Those lists are the briefs. The more specific the frustration, the better the tool.
4. Get interviewed by the agent first
Why does this interview step matter? Here’s what Louise said:
“It’s easy to get stuck in prompt loops improving the UI of your app, and making constant incremental improvements, rather than making sure the app achieves its overarching goal. This leads to a lot of token waste. Instead it helps to plan what you want beforehand and spend time talking/being interviewed by the Agent before you start building.”
Again, full honesty: I didn’t do this myself. But it’s such a great idea. The next time we run a hackathon, or even just me building something for myself, I’m going to do this.
You should too.
5. End the week with demos
Everyone shows what they built, why, and how it works.
The demos are where the cross-pollination happens, where someone realises their tool would be 10x better with the data another teammate’s tool produces, and where the next week’s work plans itself.
And, naturally: build it in Agent A. (Yes, I’d say that. But the shared workspace is the difference between “everyone has a folder of one-off ChatGPT chats” and “the team has a library of working tools that keep working next week”. The hackathon is the spark; the workspace is what keeps the lights on.)
Final thoughts
The marketers winning with AI right now are not the ones with the cleverest prompts or the longest stack. They’re the ones who took a week to look honestly at their own work, picked the boring repetitive parts, and built the small tool that handles them.
Stop trying to “use AI more”. Start by listing the five things you keep doing manually that you shouldn’t have to.
