Prompting Is Dead. The Skill That Replaced It Is Delegation.
How I went from spending $1,000 a month on an AI agent that broke every other day to running three assistants across seven businesses for $20.
Most people are still using AI like a faster Google. You type a question, you get an answer, the conversation ends. Sure, the answer helped. But then you go and do all the actual work yourself anyway.
That was me for a long time, too. And I want to tell you what changed, because it is not what the prompt-pack people are selling you.
How I Actually Got Here
In January, a tool called OpenClaw hit the market. It was one of the first autonomous AI agents that regular people could actually download and run, and it went bananas on GitHub almost overnight. People started running it on a dedicated Mac mini so the agent never touched the computer where they kept their bank logins. That created a global shortage of Mac minis. We live in Bali, there were none on the island, and the last two in the entire country were sitting in Jakarta. My husband flew there to grab them.
Then I went to work. I spent about a week building a 42-page security and hardening document because I was not about to hand an autonomous agent the keys without knowing exactly what it could and couldn’t do. We launched in January, and February and March were me deep in the thing, learning what these agents could actually do once they had memory and tools and workflows.
It was rubbish, honestly. It broke constantly.
Every time the underlying model updated, my workflows snapped, so I’d fix something on a Tuesday and watch it break again by Thursday. And because you couldn’t connect your existing subscription yet, I was paying per API call. About $1,000 a month for an agent that worked maybe half the time.
Then Hermes launched in April. It was self-learning and self-improving, and it actually showed you what it was doing behind the scenes. Around the same time, the guy who built OpenClaw got hired by OpenAI, ChatGPT opened up its subscriptions for agent use, and they shipped their top model. I connected my Hermes agent to a $20 ChatGPT plan and my mind was blown.
Now I run three agents across seven businesses on that same $20 plan, and I have never once hit my limit. The expensive, breaking-every-other-day version of this is over.
An Assistant Is Not a Chatbot
People hear “AI assistant” and picture a smarter chatbot. It is not the same category of thing. Here is what separates the two, and why I keep calling mine employees.
Memory. It knows my offers, my voice, my audience, and my standing decisions. I never re-explain my business at the start of a conversation.
File access. It reads my documents, my spreadsheets, my messy exports. I can hand it a 10-page PDF and ask for a summary, and it does it.
Tools. It can act, not just talk. It searches, checks things, creates files, and touches my actual systems, within the limits I set.
Recurring tasks. It does jobs on a schedule without being asked. This is where it stops being a chat and starts being staff.
Workflows. Chains of steps it runs the same reliable way every time, like a documented process it executes on its own.
Stack all five layers and you have something that takes the work off your plate and runs it without you standing over it. Mine draft my content, chase my follow-ups, and brief me before I’m even awake.
The Skill Is Delegation, Not Prompting
Everyone is still trying to sell you prompt packs. Learn the magic words, get the better answer. Prompting was a real skill in 2023, 2024, even 2025. We are in a different world now.
The skill that matters from here is delegation. It is the same muscle you use to manage a human assistant. You brief it once, clearly, then it goes and runs the task on repeat without you re-briefing it every single time.
Which is exactly why non-technical people are often BETTER at this than engineers. If you have ever managed a team, an employee, a household, or three kids and a husband, you already know how to instruct someone, set boundaries, and hand off responsibility. You already have the skill. You just have a $20-a-month assistant to hand it off to now.
The Five Workflows I’d Build First
If I were setting up a brand-new assistant tomorrow, this is the order I’d build it in. Each one learns a job, and each one replaces something that is currently eating your time or leaking your money.
The capture catcher. One place to dump everything. Ideas, reminders, voice notes, the “don’t let me forget this” thoughts. The agent sorts the mess into follow-ups, content ideas, tasks, and decisions. The job it’s learning is catch what I drop. What it replaces is your 70 sticky notes and the 3 a.m. things you type into your phone and never look at again.
The follow-up keeper. Once it’s catching things, you hand it the open loops. The proposal you sent that needs a nudge in 72 hours. The person who said circle back to me. It creates the reminders and surfaces them when you need them, so the job it’s learning is help me not drop balls. For service, network, and affiliate folks, this one is your gem, because the money you leave on the table is almost always sitting in a follow-up you forgot to send. I also use it for regular life things like reminding me to follow up with my adult kids about something, or reminding me to order the shower filters when I’m back at my desk.
The morning briefing. This one changed my whole day. Once you’ve done a real brain dump and the capture catcher is running, your agent has enough context to brief you. What’s overdue, what came in overnight, what needs a decision today. The job it’s learning is start my day by telling me what matters, and what it replaces is the first 45 minutes you waste just orienting yourself. Mine also pulls the top news stories in my industry every morning, which keeps me current and feeds half my blog ideas.
Research and decisions. Hand it the question instead of opening 40 tabs. Compare two platforms for my use case. Read this contract and tell me where it commits me and where it limits me. Pressure-test this launch idea and tell me where it breaks. It comes back with a brief, not a pile of links for you to wade through. The job it’s learning is help me think, not just remember, and what it replaces is the afternoons that vanish down a research rabbit hole.
The content multiplier. Once it has your voice, your offers, and your audience, you put one idea in and pull all the pieces out. Here is my real example. I post five Substack notes a day, every day. My agent pulls my last two blogs once a week, writes 35 notes in the format we already agreed on, and queues them. I spend 30 minutes on a Tuesday reviewing them because I’m still a control freak about my own voice. That used to be three to four hours a week. The content creation engine runs itself now, and I just approve.
There’s a bonus one I’d add the moment you trust it: inbox triage. I have about six email accounts and 40 to 50 emails landing overnight. I ask one agent what actually needs my attention, and she tells me. She can read and report. She cannot send. That single boundary saves me from getting sucked into my inbox and losing an hour I was supposed to spend somewhere else.
Boundaries: An Agent Is As Safe As You Make It
Your agent is exactly as safe as the boundaries you give it. Same as a new hire. You don’t hand a brand-new assistant your bank logins on day one, and you don’t hand them to an agent either.
In practice that looks like a few simple things. The agent has to ask before it touches something new, and you get a popup to allow once, always allow, or never. You keep anything sensitive isolated, which is the whole reason people run agents on a separate Mac mini instead of the laptop where they’re logged into everything. And when several workflows could touch the same task, you give each one its lane so they don’t collide.
For me, none of my agents touch money. They don’t send a single email or message on my behalf. I review every draft before it goes out, and because I write in my own voice with a 98-point guardrails document, I edit almost everything by hand. An article still takes me about an hour of manual editing after the first draft. That is on purpose. Done well, an assistant gives you back hours. Get it wrong, and it ships something in your name that you never approved. The authority expands as it earns trust. You steward it the same way you’d steward your money or your people.
Starting Is Easier Than the Setup Sounds
This is not scary to set up. You need a computer, a $20 ChatGPT subscription, and one install command that walks you through a short setup wizard. Then you write your first briefing.
There is a little tech setup at the front, but the agent helps you do all of it. When I connected mine to my Gmail and calendar, I didn’t even know what an OAuth was. I asked the agent how to do it, it walked me through every step, and I was connected and tested in about 30 minutes. If you can send a text message, you can do this. If you can leave a voice note, you can do this.
The Gap Used to Be Technical. It Isn’t Anymore.
Six months from now there will be two kinds of people. The ones with an assistant running their follow-ups, sending their briefings, and fueling their content, and the ones still typing questions into ChatGPT like it’s a fancier search bar. That gap is going to keep widening.
Here’s what’s different about right now. The thing that used to separate those two groups was technical skill. In January, this was genuinely hard, and I have the $1,000 invoices to prove it. It is not hard anymore. The only gap left is whether you set it up and start using it. I’ve handed you the map. Go build the thing.
I recorded a whole 2-hour masterclass on this last week, and the replay is available now for $24. You can get it here: https://agent.aiwithleah.com/
And if you want the step-by-step on how to set up your agent, make sure to also grab the $12 Quick Start Guide.