The best AI users aren’t better at prompting. They’re better at managing.
Six months ago, I developed a workflow to solve a software problem. AI coding assistants were fast and capable, but they kept losing the plot — forgetting decisions, contradicting themselves, rebuilding things that already existed. The issue wasn’t intelligence. It was organisation.
So I started structuring my AI work the way I’d structure a small team: one agent held the big picture, others were brought in for specific tasks, and shared documents kept everyone aligned. I called it the supervisor/specialist pattern.
It solved the coding problem. Then I noticed it solved several other things too.
The pattern
The idea is straightforward, though it takes a bit of discipline to stick with.
One AI conversation acts as an AI supervisor. It holds the strategic context — what you’re trying to achieve, what’s been decided, what’s changed along the way. It never does the hands-on work. It thinks, plans, and prepares briefs for the agents that do.
Separate conversations handle the execution: draft this proposal, analyse these numbers, research these competitors. Each one gets a focused task and just enough context to do it well. When it’s done, the output goes back to the supervisor, which tracks progress and sets up the next piece of work.
The specialist doesn’t need long-term memory — it just needs a clear brief and a defined outcome. The supervisor doesn’t need to do the work — it needs to maintain the thread that connects one task to the next.
If that sounds familiar, it should. It’s how effective delegation has always worked. The AI just makes the pattern more visible.
Why this matters beyond software
I originally built this for code, but the underlying problem — that context degrades over time, and AI doesn’t manage its own continuity — turns up everywhere.
Think about a head of marketing running a brand refresh. Over several weeks, she uses AI to develop positioning, draft messaging, and write copy. Each conversation starts fresh. By week three, the copy has started drifting from the positioning the AI itself helped develop in week one. Not because the AI has done anything wrong — it simply has no memory of its own earlier thinking.
Or a consultant preparing a pitch. The market research, competitive analysis, proposal, and financial model are all produced with AI across different sessions. The consultant is the only thread connecting them — and they’re too busy to notice when the proposal quietly contradicts the research.
In every case, the human is already doing the supervisor’s job. They’re just doing it in their head, inconsistently, without the discipline of writing it down. The pattern simply makes that invisible work explicit.
The system that holds it together
Over time, I’ve settled on four practices that keep the whole thing working. None of them require technical skill — just a willingness to be organised about it.
The orientation document. Every project I run — whether it’s a software build, a marketing strategy, or a music production — has a single file that contains every key fact an agent needs to get started: what the project is, where the important files live, how things get published, and an index to all the reference documents. Think of it as the induction pack for your virtual team. Every agent reads it before doing anything else.
The reference library. Behind that orientation document sits a structured set of plans, analyses, resources, and reference files — and here’s the key part — maintained by the agents themselves. One of the most valuable is a naming reference: every brand colour, product name, and key definition recorded in one place. When an agent creates something new, it checks the reference first. No guessing, no drift, no two agents using different names for the same thing.
The wrap-up discipline. After every significant piece of work, the agent updates the relevant documents — the orientation file, the references, whatever has changed. This isn’t optional. The AI’s last job on every task is to keep the shared knowledge current for whoever picks up the thread next.
The reality check. Every so often, I run what I call a system state confirmation — essentially testing the documentation against what actually exists. Does the project plan still match the code? Do the brand references match what’s been published? Has the orientation document drifted from reality? This is what I’ve written about elsewhere as the AI amnesia problem — and this is the discipline that catches it before it causes real damage.
Is there overhead in all of this? Honestly, yes. But that’s what management is, isn’t it? Recording things systematically, keeping people informed to the level they need, and constantly testing your internal picture against reality. The only difference is that the team happens to be artificial.
This article is an example
I should mention — I’m using the pattern right now, as I write this.
I work with one AI conversation about my publishing strategy: what I’m writing, why it matters, who it’s for, how the pieces connect to each other. This supervisor doesn’t write the articles. Together, we plan them and assign tasks to a separate specialist chat agent.
So, when I’m happy with the strategy, I open a separate conversation, give it the orientation document and the relevant context, and let it work. This is normally my own article draft, or a set of bullet points with the facts that need to be in the article. Who the audience is. What I’m trying to achieve with the publication. Then the specialist agent produces a draft, which I critique and may pass through other AI agents for checking.
When we’re done, the key decisions and any new thinking feed back into the reference documents for next time.
That’s human-centric AI in practice. I’m not waiting for the AI to figure out what matters. I’m telling it — and structuring the work so that telling it once is enough.
The management skill nobody’s teaching
There’s an irony in how most organisations approach AI adoption. They invest in prompt engineering workshops and tool training. They teach people how to talk to AI.
What they don’t teach is how to manage AI — how to structure work across multiple conversations, maintain continuity over time, and build the kind of contextual infrastructure that makes AI output reliable rather than lucky.
The people getting the most from AI aren’t the ones writing the best prompts. They’re the ones who’ve figured out that managing an AI workflow is really just managing — applied to a new kind of team member that happens to be extraordinarily fast, deeply knowledgeable, and completely incapable of remembering yesterday.
Once you make that shift, the AI stops being a clever tool you occasionally consult and starts being a managed capability you can genuinely rely on.
That’s the difference between using AI and actually working with it.
