Tell me what you want: what an advanced Claude Code workshop taught me

In an advanced Claude Code workshop run by Boon Kgim this July, the instructor put up a diagram. One orchestrator giving direction, several sub-agents underneath it, each one handling a single narrow task and nothing else. I recognised it immediately. Not from a course or a book. From my own kitchen table, months earlier, rebuilding JobPilot.

A pattern I built without a name for it

I have written elsewhere about how JobPilot became five sub-agents instead of one. A search agent, a scoring agent, a resume agent, a cover letter agent, a logging agent, each doing a single job instead of one agent carrying the whole process. I did not arrive at that structure from any technical framework. I arrived at it because I know my own job search process cold, and once I could see it clearly on the page, the seams between the separate decisions were obvious.

At the time, I had no name for what I had built. It was just the shape the problem took once I understood it well enough.

What the workshop actually gave me

Sitting in Boon Kgim’s workshop, I did not learn a new trick. I learned the reasoning behind one I had already used. The term for it is context engineering, which sounds more complicated than it is. It simply means being deliberate about what information an AI agent is carrying at any given moment, rather than handing it everything and hoping it picks out the relevant parts.

That is exactly what was wrong with JobPilot’s first version, and exactly what the five sub-agent structure fixed without me having a word for it. Each sub-agent now holds only what it needs for its own task. Nothing more. The workshop also covered building reusable skills, which in plain terms means writing down a process once, clearly enough that an agent can follow it every time, instead of re-explaining it from scratch in every conversation. Again, a formal version of something I had already done the hard way, one exhausting explanation at a time.

Tell it what you want, what you really really want

The heuristic I took away, and the one I would hand to anyone building an agent for the first time, comes down to two things. Before you hand a complex task to an agent, be painfully specific about what you actually want out of it. Not roughly specific. Specific the way you would be if you were briefing a new hire on their first day and would not be in the room to answer follow up questions. It is worth remembering the Spice Girls had this right before any of us did. Tell it what you want, what you really really want.

The second thing is to stop asking one agent to hold an entire process in its head. If two parts of a workflow are genuinely separate decisions, decisions that do not need to see each other’s reasoning to be made well, give them separate agents. That is the difference between an agent that starts guessing and one that stays sharp.

Why the validation mattered more than the vocabulary

What stayed with me after the workshop was not the terminology. It was the confirmation that this was not a lucky accident specific to my job search. It is a repeatable way of thinking about any complex process, which means it is something I can actually teach, not just something I happened to stumble into once.

That distinction matters more than it sounds. A trick that worked once is an anecdote. A pattern you can name, explain, and apply to someone else’s workflow, whether that is a hiring process, a client onboarding sequence, or a shop floor reporting routine, is something closer to a method.

Where this leaves you

If you are building your own first agent, start with the two things above. Get painfully clear on what you want before you hand over the task, and give separate decisions separate agents rather than asking one to do everything. You do not need the vocabulary to get this right. I did not have it either, the first time.

If you are further along, running a business or a team and wondering whether this way of working could hold up somewhere more complicated than a job search, that is the conversation I would rather have directly. Get in touch and we can look at your process together.