Every week someone posts about making money with AI — fast, easy, almost automatically. Get rich with AI. I didn’t buy it. I tested it.
This is the honest account.
The Experiment: I Ran the Amazon Affiliate Playbook
When I started with AI seriously, I wasn’t looking for shortcuts. I was looking for something that worked. The most talked-about template was clear: Amazon affiliate products pushed through Pinterest, with AI handling the content side. Simple pitch — set it up, let it run.
I ran the experiment properly. Here’s what the data showed.
The first obstacle was product images. Plain white-background shots, the kind you pull straight from manufacturer listings, perform badly on visual platforms. Lifestyle and context images get roughly three to four times more engagement than isolated product shots. That’s not a small gap. It’s the difference between content that stops the scroll and content that gets skipped.
So I tested AI image tools to close that gap automatically. In practice:
- One tool couldn’t isolate the product from the source photo at all. It grabbed something similar-looking instead of the actual item.
- Others placed the product in the wrong dimensions, or dropped it into an environment that had nothing to do with what it was.
Better models exist. But not on the free and cheap tiers a beginner actually starts with. Getting from “product photo” to “believable lifestyle pin” doesn’t work reliably down there. And even if you solve the image problem, you need thousands of pins before any revenue shows up. The investment-to-revenue ratio — in time, in tools, in cleanup — doesn’t make sense for a solo founder without a budget. I looked at the numbers and moved on.
The Drift Problem I Didn’t See Coming
The Pinterest experiment was the visible result. The deeper problem took longer to surface.
I’d started building real automation in parallel: a server, workflows, scheduled tasks. Early sessions went well. Things connected. Workflows ran. Then, slowly, I started losing ground.
Each new session picked up close to where I’d left off, but never quite. Decisions I’d made last week weren’t in context anymore. I’d re-explain the same setup. Files piled up. At some point I was spending the first twenty minutes of every session just re-establishing what we’d already agreed on.
This is AI project drift, and it’s exactly what people mean when they search “why does my AI automation fall apart” or “Claude forgets context on long project.” The AI isn’t broken. Large language models don’t carry memory across conversations. They work with whatever is in the current context window. If your briefing is incomplete or inconsistent, they build on incomplete information.
The fix isn’t a smarter model. It’s a better system around the model. And keeping that system efficient — lean token usage, structured context, no redundant re-briefing — is what makes the difference when you’re building solo on a limited budget.
The Infrastructure Was Solid. The Structure Wasn’t.
The stack we landed on — a reverse proxy for TLS and routing, a self-hosted workflow automation tool, a Postgres database for state — is genuinely practical, and it runs at a low monthly cost. We still run it today.
But technical capability without project discipline is just expensive tinkering. We even hit a specific failure that cost half an hour to diagnose: a config update that appeared to succeed but changed nothing, because of how containerized services track file identity at the OS level. That kind of failure repeats until you know its name.
The missing layer was never better tooling. It was structure: a way to keep context intact across sessions, track decisions and the reasons behind them, and stop work from drifting further from the goal every time a chat window closed.
What We Built — and Why You Don’t Need to Wait
We didn’t set out to build a framework. We set out to make our own project stop falling apart. The drift, the lost context, the decisions we kept re-making from scratch — that was a wall we walked into again and again, and the only way through was to fix the foundation underneath the work. The framework wasn’t a side project to free up time for something else. It was the logical consequence of solving our own problem.
CoveLab Foundation is what came out of that: session hooks that inject the current project state automatically, named work steps, a live state document, and validation gates that catch drift before it compounds. We run it in production ourselves. If you keep re-explaining context that should already be set, or redoing work you thought was finished, the current version handles that now.
For teams running multi-agent setups, where different AI agents handle different parts of a project without stepping on each other, we’ve documented how those coordination patterns actually work.
Here’s the part no one posts about: the idea has to come from you. AI is the tool. The problem worth solving, the automation that fits your specific situation — that thinking can’t come from the model. “Just tell the AI what you want” skips the part that matters most. It’s where most people, including me early on, lose months.
We keep improving it. We’ve already found further limitations and built fixes for them: edge cases that don’t scale cleanly past a solo founder, patterns that still need tighter automation. That work is Foundation V2. But you don’t need to wait for it. The current version solves the drift problem today, and every improvement we ship makes it more capable. Building on a working foundation beats waiting for a perfect one.
Start with the foundation. Not the pipeline, not the product ideas — the foundation first. Everything else needs something solid to stand on.
Re-explaining context every session? Redoing work that should already be done? CoveLab Foundation is the anti-drift stack, built from the experience above.