It's the way you're talking to them. Prompt-engineering tricks stopped working two model generations ago. What actually gets AI to do real work is coaching — the same move you already run with your leadership team, applied to the agent. Under an hour of reading, five specific moves you can run tomorrow.
You already know the pattern:
You give the agent a real job. It goes sideways. You rephrase. Sideways again. Five rounds later you close the window and think "this stuff isn't ready yet."
It's ready. You just got taught the wrong sport.
The prompt-engineering industry — the Twitter threads, the $997 courses, the "10 magic prompts" PDFs — has spent three years training you to talk to agents like they're a search box. Type the right incantation, get the right answer. That worked in 2023. It hasn't worked since GPT-4.
A modern agent is not a search box. It's a direct report you hired an hour ago. It has real capability, real judgment, and zero context on your business. Prompt engineering doesn't work on a direct report. Coaching does.
Same tools, same model, same everything — brilliant work for one business owner, garbage for the next. The gap isn't intelligence. It isn't budget. It's how the owner briefs the agent, catches it when it drifts, and decides what's worth another round vs. a fresh start.
The business owners running $1–50M who get real work out of agents don't have secret prompts. They brief the agent the way they'd brief a new VP on day one: outcome, what good looks like, what to skip, who this is for. Then they walk away. When it comes back off-target, they name the specific pattern that fired and cut the loop in three seconds.
Five short modules. Under an hour. Every one leaves you with a specific move you can run tomorrow.
I’m Brad Hart. I’ve spent the last decade running masterminds for owners of $1–50M businesses — healthcare staffing shops in Georgia, e‑com brands in Utah, construction companies in Ohio, hedge funds in New York. Different industries. Same conversation on every call: “How do I get this AI to actually do the work?”
I’m not a prompt-engineering guru. I’m a business owner using the same tools you are, hitting the same walls, and writing down what shipped. The five moves in this curriculum are the ones I run every day on my own agents. This site was built by one of them.
You don’t need to buy anything to use this. If it lands and you want the room where owners run this at scale, Optimus exists for that. The curriculum stands on its own either way.
This whole site — the landing page you’re on, the subscribe endpoint, the five-day email drip, the portal collection, all five module videos, the schema markup — was built by an agent running the five moves in this curriculum. It fired every fifteen minutes on a self-driving loop. When it hit a wall, the state file said exactly what it needed. When the last box was checked, it killed its own schedule and sent me the final report.
Total human input: writing the state file, unblocking one credential, and clicking “approve” on the domain purchase. That’s it. Module 5 walks through the exact state file it ran, step by step, so you can lift the pattern for your own missions.
Every day you keep fighting your agent is a day it’s easier to just do the work yourself.
Which is the trap. You end up back where you started — the bottleneck of your own business, paying for tools you can’t get to work, watching a leaner team pass you with half the head count and double the output.
The models keep getting better. The gap between owners who coach their agents and owners who fight with them widens every quarter. You don’t need to be first. You do need to be in.
The move isn’t better prompts. It’s a different job description for the agent. That’s what this curriculum is.
Five short modules. One per day if you're going fast, or over a week if you're going deep. Each one leaves you with something specific you can do differently by tomorrow.
/fuckery · the catalog of agent bullshit
It isn't a machine glitch. It's a briefing gap. The same eight patterns show up every time — spinning on a fake constraint, inventing specifics, burying decisions in a wall of text, waiting for permission you already gave. Learn to spot them in three seconds and the sideways loop stops.
Win:You stop rephrasing. You start recognizing.
The four-part prompt structure
An agent is a direct report you hired an hour ago. Give it what you'd give a new VP on day one: the outcome, what good looks like, what to skip, who this is for. Come back with a draft. Iterate. Nothing more.
Win:One brief in. A draft you can actually use out.
/playtest · /debug · /monitor — signal, not alarm
Sessions get poisoned. Context runs out. The agent starts confidently making things up. Knowing when to kill the session and restart clean beats hammering for one more round. Here's the two-minute triage that tells you which move is right.
Win:You stop wasting an hour hoping. You know when to hit reset.
Two models · one mission · decide by output
You're doing the work anyway. Give the same brief + skills + tools to two different models running in parallel and let the output settle the argument. Frontier for the hard problem, cheaper for the routine one — the model tier IS the tool selection.
Win:Two agents worked. You picked the winner. Everyone else is still on defaults.
/pcp — persistently cron prompt
For the missions that outlive one context window: give the agent a state file, a done condition, a cron. It re-prompts itself, updates state, and kills its own cron when the job is done. Real projects that would spin for a week collapse into an overnight run.
Win:The mission ships while you sleep. You wake up to a commit log and a final report.
Keep arguing with your agent. Watch the leaner team pull further ahead. Six months from now nothing has changed and you're still the bottleneck.
Spend a year of nights and weekends learning the same five moves the hard way. Ship two bad versions before you land the third that actually works.
Drop your email. Module 1 lands in the next couple of minutes. Under an hour of reading total. Every module leaves you with a move you can run tomorrow.
Borrow my confidence for a week. Go earn the evidence yourself. That's the whole loop.