I’ve been talking to a lot of people lately who say some version of the same thing: “I know I should be doing more with AI, but I don’t know how to start.” Each week, it seems there are new models, new tools, new patterns, new takes on Twitter. It’s a moving train, and figuring out how to jump on can be daunting.
Here’s the on-ramp to building with AI that I’ve found that actually works: build something you really, really don’t like.
Not something impressive. Not something demo-able. Something that bugs you on a Tuesday afternoon and that you wish would just quietly go away.
Start with what annoys you
The first question isn’t “what should I build with AI?” It’s “what part of my job do I keep putting off?”
A few that have come up for me and people I work with:
- The daily or weekly digest nobody wants to write, but everyone reads.
- Slide decks generated for an audience of three people.
- Routing meeting transcripts and notes across the right projects, people, and follow-ups.
- The same tedious data cleanup before a recurring report.
The reason “build something you dislike” works as a starting heuristic is that you already know what good looks like. You know the inputs. You know the desired output. You know what makes the current version painful. That’s a real spec, not a hypothetical one, and if the AI version is bad, you’ll notice because you’re the customer.
Pick a tool, then stop researching
The other trap is tool research. There are too many options, and any choice you make will be partially obsolete in three months. That’s fine.
Claude Cowork/Code is a good place to start if you have access to it, but really, any agent will do. The specific things to look for:
- Repeatable prompts. Whatever lets you write a prompt once and reuse it without copy-pasting from a Note.
- Scheduled tasks. Something that can run on a cadence so you can set it and forget it.
- Connections or MCP servers. Hooks into the systems where your work actually lives — email, Slack, Google Drive, your calendar, your project tracker.
Pick whatever you already have a login for. The point is to build the thing, not to optimize the stack.
A method that fits on a sticky note
Once you’ve named the annoyance and picked a tool, the loop is short:
- Describe the problem in plain language. What is the thing? Why is it annoying? Who is it for, even if that “who” is just you?
- Imagine the automated version. What would have to be true for this to feel less awful? Don’t worry about feasibility yet – just describe it.
- Define the connections. What information or tools would need to talk to each other for the annoyance to go away? How does the information flow?
- Bundle that into a starting prompt. Hand it to your AI tool of choice and ask: help me make a repeatable, automated process for this.
The Product Managers out there will recognize the first three steps as just a mini-product brief. That last step is easy to miss. You don’t have to describe the entire system in your head before starting to build the prompt. The model is good at this part: let it propose a structure, then push back on it, make corrections, and gut-check the output.
Run it, then debrief
Give it a real trial run on real work. Not a toy version. The whole point is that you already know what good looks like, so the test is honest.
Then do two debriefs.
Ask yourself: What worked? What didn’t? Was the output something I’d actually use, or was it close-but-not-quite in the way that costs more to fix than to redo?
Then ask the AI the same questions: Where did you have to guess? What was unclear in the prompt? What would you change about the process for next time? Models are surprisingly good at critiquing their own runs when you ask directly, and the answers usually give you a sharper second draft for free.
Fold that back into the prompt, and you’ll keep those improvements for next time. A few cycles of that and you’ll have something truly useful and, more importantly, you’ll have the muscle memory for building the next one.
Why dislike is the right starting point
Ambition is a bad place to start because it’s vague. “Building something cool with AI” gives you no edges to grab onto. Annoyance gives you edges. You know exactly when the annoyance is gone.
So, what’s the most annoying part of your job?