the questions you're
already typing
long answers to the stuff everyone asks after the class.
You don't need an agent. You need your hours back.
If all you do is ask questions, a chat subscription is enough and you can stop reading. The test is simple: the moment you do the same job twice · same caption format, same weekly numbers pull, same "resize this for four platforms" · an agent pays for itself, because a job you did twice is a job you'll do fifty times.
My clipping pipeline started as one prompt. It now runs a job that used to eat hours every week. That's the entire pitch. Not intelligence · hours.
Nobody needs 100 employees. But 100 alarm clocks? That's just... a normal amount of alarm clocks.
The clickbait number hides a boring truth: an "agent" in a fleet is usually one small routine · check this, post that, file this report. Three real reasons you end up with many:
- Roles. The scout, the poster, the editor, the analyst. Separate jobs, separate context, separate rules.
- Permissions. The one that reads trending pages has zero access to your accounts. The one that posts can't touch your files. Small jobs = small blast radius when something goes wrong.
- Parallelism. Forty files to convert is one agent forty times or forty agents once.
You don't hire a staff. You set timers with rules attached.
Treat it exactly like hiring a contractor: least access first, more as it earns trust.
- Read-only before write. Let it look before it can touch. Most tools have permission modes · use them.
- Approval gates on anything public. Posting, sending, publishing: it drafts, you approve, until its hit rate proves out.
- Keys never go in chats. API keys live in env files or a secret manager. If you ever pasted a key into a chat: rotate the key, then delete the chat. Same for files you uploaded.
- Never ship keys in your builds. If your site or app calls an API, the key lives server-side (environment variables on the host), never in the frontend code · anything in the page source is public to everyone who opens dev tools.
- One key per project. So you can kill one without killing everything.
- Prefer OAuth over pasted tokens. A browser-login connection means you never handle the secret at all, and it refreshes itself. Pasted tokens expire silently and end up in config files you forget about. Ask me how I know.
- The browser trick beats keys entirely. An agent that drives your already-logged-in browser needs zero API keys for that job. Nothing to leak.
- Screensharing? Know what's on screen before you hit share. Env files, token settings pages, password managers. (I checked mine before this class. Twice.)
- Vet what you install. Skills, plugins and connectors run with real permissions. Install from sources you trust, skim what they actually execute, and take the risk scanners seriously when they flag something.
The meta-answer that feels like cheating: ask the agent.
find the official MCP connector for [tool]. walk me through connecting it step by step like i'm dumb. before we use it, tell me what permissions it asks for and test it with one read-only action first.
- Your AI app has a connector directory built in · start there before hunting the web.
- Official beats community when the connector touches accounts or money.
- First action after connecting: something read-only. Confirm it works and see what it can see, before it can change anything.
This is the workflow I actually run, taught as a pipeline. Every stage is hour 1 or hour 2 of the class wearing work clothes.
- Define the offer. Let the agent interview you: what do you actually sell, to who, and what does "done" look like. Most people skip this and automate noise.
- Price it with receipts. Research what competitors charge · always with cited sources, never vibes. If you don't know your rate, the market does.
- Scout on a cron. The agent searches for qualifying signals on a schedule: funding raises, hiring posts, product launches, public complaints about a problem you solve. Signals that say "has money" AND "has the problem."
- Qualify and score. Everything lands in a list with a why: what the signal was, why they can pay, why you fit. You review the list, not the internet.
- Open with something made for them. Not a template blast. A branded one-pager, a graphic, even a meme built from THEIR recent news · that's the hour 1 skillset pointed at sales. Relevant + recent + visual beats "quick question" every time.
- Pipeline the follow-up. Track it somewhere the agent can read and write (a sheet, Notion, any free CRM). It drafts follow-ups, crosses off dead ones, and files a weekly report on what's moving.
The honest rule: automate the sourcing, the research and the drafting. Keep the sending human until the drafts prove they don't need you · and check the platform's rules on outreach before you automate any of it.
you know what i sell: [offer]. every [monday]: search for [funding raises / hiring posts / launches / complaints] in [niche]. score each lead: the signal, why they can pay, why we fit. add to the pipeline sheet with a one-line opener idea based on something recent and specific to them. draft, don't send. friday: report what moved, what died, and one thing to change about how you scout.
Reading that holds up for offers and outreach psychology: Alex Hormozi's $100M Offers and $100M Leads. Steal the frameworks, run them through your agent.