• MediaLauncher
  • Posts
  • Building an APP with AI as a MARKETER: Speed, Debt, why it DOESN'T WORK as they advertise it, and WHAT TO DO INSTEAD

Building an APP with AI as a MARKETER: Speed, Debt, why it DOESN'T WORK as they advertise it, and WHAT TO DO INSTEAD

My experience, a real report from the trenches. What AI helps with, where it collapses, and how to start small without sinking.

Author

Francesco Cifardi
October 29, 2025

It’s been almost three months since my last update. The silence wasn’t planned, it just happened as I went heads-down, building MediaLauncher.

For new subscribers, it's my attempt to build an AI-powered version of the Google Ads Editor, a centralized workspace where advertisers can manage campaigns across platforms (Google, Meta, and other Ads platforms) using natural language. 

Clarifier: I use AI in two ways: (1) As a coding copilot, and (2) Inside the product (natural-language edits). This post is about #1.

The vision is big. The current reality? We’re still in private alpha.

One post that recently caught my eye was from Bob Meijer, who shared that he built a full RSA testing framework and an AI-driven app in just a few hours. 

Posts like that are inspiring. They make it look like we’re entering a new golden age of building software, where as they say, “you can just do things.”

And honestly that’s exactly the energy that pushed me to start building too.

The honeymoon phase

As you may know, I can code in JavaScript. That’s why I was quite successful at publishing a bunch of Google Ads scripts in the past 2 years, but I’m not a professional developer, so I never tried to build a full-blown web app.

I’ve been observing the progress of LLMs ever since I started publishing my automations, and it’s needless to say that I was completely sold on the promise of building a SaaS, without being an engineer by trade.

The catalyst moment was a terrible experience with a client (who scammed multiple vendors - may publish this story one day) and so I thought to myself: “that’s it, I’m gonna build a software, so I don’t have to deal with bad clients every again”. 😅

I started with Lovable, switched to Cursor, and eventually landed on Claude Code. 

It felt magical: I could describe what I wanted, and in minutes there was working code, even with my limited web development knowledge. I thought, This is it, the future of building software.

As a marketer who always wanted to be more technical, it was intoxicating. The ability to bring ideas to life without relying on developers felt liberating.

The reality check

When you use AI for building software, you quickly notice that it handles isolated, well-scoped problems (like Google Ads Scripts) beautifully. But as soon as your project involves interdependent systems, state management, or multi-file logic, the illusion of competence cracks. 

An LLM doesn’t understand systems, it predicts text.

The more the app evolved, the more unpredictable the AI became. It forgot earlier code, introduced inconsistencies, and refactored things that didn’t need refactoring. State management in particular became a nightmare: overlapping systems, untraceable logic, and failed refactors that created more mess than they solved.

I realized I wasn’t building faster anymore. I was just accumulating technical debt faster.

The reason for this is complexity.

Unless you TRAIN an LLM on your codebase, it’s not going to retain all the particular information about the app’s systems in its memory.

As of right now, you are limited to “claude.md” files + your prompt to provide the model with Context. This is essentially a sort of short-term memory that resets evey time you start a new session (chat).

The size of this “context window” is currently the main bottleneck for AI when it comes to tackling increasingly complex tasks.

So, what needs to happen, is that YOU have to breakdown tasks (features) into bite-sized manageable chunks that the model is able to handle.

For complex requests, this is just impossible for you to do, if you don’t know how to software and code works at its core.

Don’t take my word for it, hear what Andrej Karpathy (founding team @ OpenAI, former Director of AI @ Tesla) has to say:

“When you’re talking about an agent, […] you should think of it almost like an employee or like an intern that you would hire to work with you.

So for example, you work with some employees here. When would you prefer to have an agent like Claude or Codex do that work? Currently, of course they can’t.

What would it take for them to be able to do that? Why don’t you do it today?

The reason you don’t do it today is because they just don’t work.

They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use and all this kind of stuff. […] They don’t have continual learning. You can’t just tell them something and they’ll remember it. And they’re just cognitively lacking and it’s just not working.

I just think that it will take about a decade to work through all of those issues.”

That’s why I called in a real developer to help clean up the mess.

I also decided to stop and go back at learning React, NextJS, and web development properly myself.

The turning point

This experience reminded me what the excitement made me forget:

AI doesn’t eliminate the need for fundamentals. It just changes how you use them.

To use AI effectively, you need to understand enough about the system you’re building to guide it, correct it, and structure your prompts around sound logic. Otherwise, you end up building a house of cards, fast.

I mean, sure, there are app ideas that are feasible (look at Alfred’s BackdropBoost) but it’s not the case for MediaLauncher. To the vast mojority of marketers who want to build software I’d say…

A note for other marketers looking to build something

If you’re a marketer seeing posts like Bob’s and thinking, “Maybe I should build something too,” the answer is yes, you should. But do it with open eyes.

AI can supercharge your creativity and speed, but it can’t replace understanding. The more you know about the foundations, how apps work, how data flows, how systems talk to each other, the more powerful AI becomes as a partner.

In fact, before working on MediaLauncher, I also built a small Chrome Extension to solve a personal pain point, and it reminded me that you don’t need to build a full SaaS to create something useful.

If you’re eager to build something, start small

I got my LinkedIn account suspended multiple times for using automation tools. So I built a personalized messages Chrome Extension that saves me time without breaking the rules. It adds small buttons inside LinkedIn chat with my most-used “welcome” templates and automatically inserts the user’s first name. Now, when 100 people connect in a day, I can greet them all manually, but in seconds.


I was able to build it with AI because its scope and complexity is small enough for LLMs to keep it in their context window.

So if you want to get started, here’s my advice: the most effective way to learn and build confidence isn’t by trying to create the next big platform, it’s by starting small.

Start with micro-projects. Things that solve a single problem well. Things like:

🧩 Chrome Extensions

Quick, lightweight tools that live in your browser and save time during campaign management.

  • A “UTM Builder” that lets you right-click any link and copy it with predefined tracking parameters.

  • A “Google Ads Asset Previewer” that shows how responsive ads render across devices.

  • A “Bulk Keyword Cleaner” that formats, deduplicates, and trims keyword lists from CSV uploads.

  • A “Performance Snippet Saver” that lets you save top-performing headlines or descriptions directly from Google Ads UI.

Powerful automation for account optimization or alerting.

  • A “Budget Pacer” that adjusts daily budgets based on monthly spend progress.

  • A “Search Term Cleaner” that automatically pauses low-quality terms or alerts you when CPC spikes.

  • A “Conversion Drop Detector” that emails you when conversions dip beyond a threshold.

  • A “RSA Asset Performance Reporter” that sends weekly summaries of asset ratings.

  • A “Competitor Overlap Monitor” that uses Auction Insights to track impression share changes.

📊 Spreadsheet & API Automations

Micro-tools that sit between data and insight.

  • A Google Sheets dashboard that pulls campaign data daily using the Google Ads API.

  • A “Creative Testing Tracker” that aggregates RSA variations and conversion metrics.

  • A “PMax Device Split Visualizer” for analyzing device-level data over time.

  • A “Daily Spend Anomaly Detector” powered by simple formulas or Apps Script.

🧠 Lightweight Apps or Dashboards

Small web apps that make internal workflows smoother.

  • A “Campaign Naming Validator” that checks if names follow your team’s convention.

  • A “Landing Page Health Checker” that detects 404s or slow-loading pages.

  • A “Meta + Google Performance Merger” that combines metrics for unified ROAS reporting.

  • A “Creative Rotation Tracker” for visualizing ad fatigue.

🧾 No-Code & Workflow Automations

Simple systems built in Airtable, Notion, or n8n that connect your marketing stack.

  • A Notion dashboard that tracks campaign launches and automates status updates via Zapier.

  • An n8n workflow that sends alerts when new search trends appear in your keyword list.

  • An Airtable CRM that integrates form submissions and lead quality scores automatically.

  • A Make.com flow that syncs ad performance into Slack for daily standups.

Every small project teaches you something new, about logic, structure, data, and how systems actually work together. It builds momentum. And, over time, that momentum compounds into the skills you need to tackle something bigger.

Closing thoughts

Agents are not ready yet, but it’s not a bubble. Big Tech companies have clearly over-invested in AI talent and research, and we’re now seeing the first major correction. Meta just laid off hundreds from its FAIR and Superintelligence Labs teams. Microsoft, Amazon, and Intel are doing the same, not because they’re abandoning AI, but because they’re restructuring around what’s actually working: infrastructure, model serving, and applied AI.

Yes, we’ll probably see more layoffs in technical and AI-related roles through 2026. But this isn’t like the dot-com crash of 1999. The difference is infrastructure. Back then, much of the internet was theoretical, today, AI already runs on billions of dollars of GPUs, data centers, and production systems built by companies with the deepest pockets in history.

In other words, there is a speculative layer “at the application level”, overhyped startups, inflated valuations, and too much capital chasing too few real problems, but beneath it lies something much more durable: structural transformation. AI has already become part of the global industrial base.

So, no, agents aren’t ready, startups are shaky, and some teams will shrink. But the technology is good enough for you to build isolated, well-scoped projects.