How I Went From Prompting to Producing

About a year ago, I did a Lunch & Learn for EM on building AI into marketing workflows. We had a grand ole time. I turned a page of meeting notes into a campaign brief. I built an Instagram ad from a bunch of mixed inputs, and I ran some market research that pulled 23 sources in about eight minutes.

All of that stuff is still pretty useful and pretty amazing. At the time I would have described AI as a collaborator, or a brainstorming partner, or sometimes a teacher (teach me how to do a bunch of stuff in Excel I never knew how to do before). Now I find it much more like having an assistant, or even a small team, that I manage.

What’s changed with AI since last year?

A lot of what was important then is still important now. You still need to give it plenty of context to get good results, though there are much better ways to do that now than typing it into a chat window every time. You still have to know how to evaluate what it gives you.

What changed is that it can reach the tools I actually work in. Claude (or ChatGPT) connects directly to programs like Airtable, Gmail and Drive, and through Zapier to most of the other tools. Super useful. Now I can just ask it to do the thing, so I’m building things instead of just getting information back.

What’s in your head?

Sometimes I use AI to get what’s in a few people’s heads into something everyone can use.

A nonprofit client had years of strong, consistent writing across annual reports and social posts, but none of it was written down. The voice was there, it just lived inside a couple of people’s heads, and every new piece started from memory.

I used AI to read their published work the way an editor would, looking for the phrases they repeat, their tone, the words they lean on and the ones they avoid. The main rule for the project was that the AI was describing their voice, not inventing one for them. Now they have a voice and brand guide their marketing team can write against. They can also load it into ChatGPT or Claude to draft in their voice, or to check whether something already written is on brand.

What do you not have time to do?

Sometimes I use AI to do something that would obviously be useful, if only anyone had the time.

A small tech company needed competitive intelligence. They couldn’t justify hiring someone for it, and nobody on their marketing team had the bandwidth to do it properly, so it kept not happening. I built them a virtual newsroom. The “reporters” are AI agents, which just means an AI that goes off and does a job instead of waiting for you in a chat window. Each one covers a beat, an area like industry news or competitor positioning, quotes its sources word for word, and links every claim, so anything in a briefing can be easily checked in about a minute.

I also gave it a memory. It knows what it reported last month, so each briefing only tells you what is new. We started with a single reporter on one beat, and they have already used what it turned up to shape some of their marketing.

Make it follow a process

Somewhere in the last year I stopped typing “build me this cool thing” into a chat window. That approach, also known as vibe coding, is kind of magical, but you don’t know how it built the thing or whether it will break after a couple of uses.

One of the biggest tools in my kit is a framework called BMAD. It’s free and open source, and it organizes AI coding agents like an actual software team, moving a project from a brainstorm through a brief and a design and into small pieces of work that get built and reviewed one at a time. It is slower at the start. It is also the difference between a demo and something I can confidently hand to a client.

For example, a residential remodeling contractor needed a way to give homeowners a ballpark budget during a first sales call. Getting to a number used to mean doing much of a real estimate, which is hours of work before you know whether the job is yours or even real. We built an Airtable price book out of his own completed jobs and accounting history, and an app on top of it that produces that ballpark from the work he has actually done.

The BMAD process really helped out in a couple ways. The first was the brainstorm, where we set out to attack our own early conclusions instead of defending them. That’s how we found out he only had about a dozen kitchens in his job history. That is far too few for a model to learn a pattern from, it would just memorize those twelve jobs, so the pricing runs on rules drawn from his own numbers instead.

The second was the testing. By the end there were 681 automated checks that re-run every time anything changes. Vibe coding does not get you 681 tests. A build without them might look fine until real users with real data start using it, and then it breaks.

Give it context (still)

AIs have limited short term memories, which is why they can get weird when a thread runs long. Everything you explained an hour ago is gone the next time you open a new chat, so you end up explaining the same project over and over.

So I built an assistant that uses a “second brain,” a folder of notes that the AI writes to and reads from between sessions. It keeps notes on my clients, my projects and my work, and it reads them at the start of every session. What we decided, what we tried, what we ruled out and why, who everyone is, etc. I don’t have to prompt it to go and find something, it’s just there. There is nothing fancy about it. It’s a folder of text files, and the work is in keeping them current, not in the tooling.

I got the idea partly from BMAD, which writes everything down as it goes. Every decision, every requirement, every bit of architecture, in plain text files on my own computer. The AI keeps referring back to those files, so it always knows what it is building and where we are in the process.

How do I know it’s any good?

If it’s building things I couldn’t have built myself, how would I know whether they’re any good?

I’m not the one doing the hands-on work, but I’ve managed web projects with people before. I know the process a good team follows, and where the review steps need to be so you end up with something solid rather than something that just looks finished.

The difference is that I can’t trust an AI the way I can a good person or a good team. It doesn’t have morals or lived experience. It just knows stuff, and knows how to do stuff. So instead of hiring someone I trust, I make the AI follow a process I trust, and that usually leads to the best outcome. That’s most of why I use BMAD.

Then I check the result myself, by using it. When we tested the contractor’s intake form for new jobs, it failed because a sharing setting was locking out anyone who didn’t have access to the database. You catch that by filling out the form like a normal person would, not by reading the setup.

Neither the process nor the check at the end tells you whether the thing was worth making in the first place. That part is taste, and I think it’s the most important part of the job now. There’s so much slop out there because cranking things out has never been easier. The hard part is knowing whether it’s any good.

A year ago, in front of the EM community, the question was what AI could do. The more useful question now is what you want it to make, and whether you’d put your name on it when it’s done.

If you have something stuck in someone’s head, or something useful that never gets done because nobody has time, I’m happy to talk about it. I can also tell you what one of these actually takes: how long, how much of your team’s time, and who keeps it running afterwards.