How to manage an AI Agent, not just ask it one-off questions

098x

Four principles that separate people who actually delegate work to AI agents from those who just keep asking one-off questions.

Most people working with AI today keep doing the same thing: they ask a question, get an answer, then ask another. And another. Simplified, they’re using artificial intelligence as a faster Google search.

Matt Pocock, a developer and content creator, named this clearly: “Skill Hell.” You might know the term from a different context as “Tutorial Hell” – endlessly learning skills that never come together into a working whole.

In his video Building Great Agent Skills: “The Missing Manual” (June 2026), Pocock explains why most people work with AI agents the wrong way, and how to fix it. The video speaks mainly to developers, but the principle holds for anyone who wants to actually delegate work to AI, not just consult it for one-off answers.

Four things that actually matter

Pocock proposes a checklist for every skill you hand to an AI agent. It works for developers, and just as well for a manager who wants an agent to handle something like a daily email review more efficiently.

Trigger – when the skill fires. Get this wrong and the agent either never uses it, or fires it at random. Both are bad.
Structure – how the skill is built inside. The steps need to be ordered and clear, because an agent handed vague instructions returns a vague result.
Steering – how you guide the agent while it works. Handing over a task and waiting for the result isn’t enough. Good instructions help the agent think and act even when it hits a snag.
Pruning – regular cleanup. Instructions age. What held true in January might not hold in June, and an agent running on stale instructions does stale work for you.

TIP: Before assigning a task to an AI agent, try answering three questions first: What exactly do I want as the output? In what format? And what should the agent not do? The more specific the prompt, the fewer corrections needed – and the better use of everyone’s time.

User-invoked vs. model-invoked

Pocock stresses this distinction, and most users never think about it at all.

User-invoked – a skill you trigger yourself. You decide when to use it, and the process stays controllable and predictable.
Model-invoked – a skill the AI model triggers on its own, deciding by itself when to run it. Tempting, but unpredictable in timing, and you’ll often end up paying for wasted compute.

Pocock prefers the first option, user-invoked. The reason is simple: you want to know what your agent is doing and why, and keep the whole process under your own control.

Why a weekend isn’t enough to figure this out

At one point Pocock makes a statement that holds true well beyond the developer world: a weekend experiment won’t prepare you to use AI and agentic work effectively.

Working with AI agents isn’t just a technical skill. It’s a new way of thinking about work: what you delegate, how you structure the task, how you check the result.

And that’s a skill you learn systematically, with the context of your own field in mind.

TIP: AI agents are just one piece of a much larger shift happening across industries and job roles right now. Want to understand the bigger picture of how AI is changing the way we work? Read our article: Smarter, faster, more efficient: AI is reshaping how we live and work.

Want to actually put AI and agents to work in your field?

Studying at EDU Effective is built on exactly this principle. The Effective MBA: Applied Artificial Intelligence and Effective MBA: Mastery in AI programs connect work with AI tools to what you already know well: your field, your practice, your everyday context.

From day one you get access to 10 modules, each split into 30 blocks of 15 minutes. Morning, evening, whenever it fits your day – 100% online.

You study without risk. If within 14 days you find microlearning isn’t for you, you get the full amount back, no questions asked, under our MoneyBack Guarantee.

Apply and find full details at EDU Effective.