When my AI screws up, it’s almost never the model.
It’s me. I’d handed over a brief that a day-one intern would’ve failed at, and I’d expected a finished job back anyway.
“Handle my inbox” is a vague request. “Sort new mail into my existing Gmail labels every hour, never send or delete anything, and flag only what needs my urgent attention” is a brief.
Same agent, very different Tuesday.
Last week covered what to build before you turn on the bot. This week is the handoff: how you actually give an AI agent a job.
Most agent failures aren’t model failures. They’re delegation failures, and you can start fixing yours today with one simple test that every manager already knows.
Run the intern test
Before you delegate anything, ask yourself one question: would I give this exact brief to a smart new intern on their first day? If the intern would come back with ten questions, the agent will too.
That’s where the slop comes from.
Here are the five questions every day-one intern asks:
- What’s the job? One task, one sentence, one outcome.
- Where’s the stuff? Which inbox, folder, doc or tool they can read.
- What does done look like? A pass/fail check, ideally with an example of good.
- What can’t I touch? Never send, publish, spend or delete without asking.
- Who checks it, and when? You review the first runs. A second set of eyes after that.
Your brief should answer all five. Most prompts don’t get past the first one.
The 10, beginner to advanced
These are delegations I actually run. Every item ends with a “done when” check, because a clear definition of done is the intern test in practice.
Beginner: read-only, easy to check
Start here. The agent reads and reports, and nothing leaves your screen without you.
- Morning brief. My agent pulls my calendar, priority email and open tasks before I’m up. Done when: it’s waiting before 7am with today’s top three priorities and a list I can delegate to agents.
- Inbox triage (never send). It sorts new mail into my existing labels hourly and sends a short digest a few times a day. Done when: every new email has a label and the digest flags only what needs me.
- Meeting notes to action items. Transcripts arrive on their own and the agent pulls out decisions and to-dos. Done when: every action item has an owner and a due date.
- Calendar conflict watch. It scans the week for overlaps and impossible travel. It once caught two Friday-morning meetings 15 minutes apart that were both going to run long. Done when: clashes are flagged ahead with a suggested fix.
- Reading digest. It reads the newsletters and X lists I don’t have time for. Done when: 200 newsletters become five things worth reading, each with a link.
Medium: needs a written brief and acceptance checks
These need real sources, a written brief and an acceptance check you’re able to verify in a few minutes.
- Prospect dossier. The agent researches the person and company before a call. Done when: it’s one page, every claim is cited, and lets me read their blog before we meet.
- Lead list from scratch. I had an agent map a B2B client’s whole target universe across five states, including opportunity they weren’t tracking, then gather the contacts. Done when: the list is deduped, every row is double sourced, and it’s loaded into the CRM.
- SEO stale-page audit. It watches the site for old stats, dead links, retired tools and stale dates. Done when: you get a ranked refresh list with the exact stale lines quoted.
Advanced: you have more than one agent, a review gate
Save these until your beginner jobs run cleanly, because at this level you’re managing a process with several moving parts, not a single assignment.
- Second-agent review. When my agents build something, it doesn’t count as done until it’s reviewed. The reviewer is never the author, and it’s a different AI (Codex reviews Claude). You wouldn’t let the intern approve their own work. Done when: there’s a recorded pass or fail with specific findings.
- An agent that learns your edits. My content agent writes down which ideas it thinks I’ll pick, and how it would edit, before it sees my choices. Then it compares. The good interns learn your taste. Done when: agreement goes up week over week.
Crawl, walk, run.
Do this today
Pick one beginner item and write its brief using the five questions, in five sentences or fewer. Run it three times with you as the reviewer, and after each run fix the brief, not the bot.
Three clean runs and it’s earned a permanent spot on your team.
Resources 📚
- Skip a Brick, Ship More Slop: what to build before you turn any agent on.
- The beginner’s guide to AI agents: which bot to start with.
Write one brief before Friday. Your agent’s only as clear as the instructions you give it.