From prompt to operating system

Some mornings, four pieces of work come back before I have finished my coffee. A market scan, a reviewed contract, a plan for next quarter, the actions from yesterday’s meeting. Each one is reasonable. Each one is waiting for me.

I started using AI to remove work. As the system became more capable, it made something else visible: my own judgment and attention had become the constraint.

At the beginning of 2026, I wrote about building my homepage with AI. Like most people, I started in a chat window, writing one-off prompts and waiting for answers. Since then, the way I work has shifted from prompting in chat to running reusable skills, built on private knowledge and company standards. Over time, it has become its own operating system, the Leadership OS as I call it: a living structure of private knowledge, reusable skills and human review.

The bottleneck moved

My first uses of AI were about production. Research became faster. First drafts appeared in minutes. Long documents could be summarised and plans challenged with much less manual effort.

The first constraint therefore moved to clarity. If AI can produce almost anything, the quality of the result depends on the quality of the question.

Then it moved again. AI could generate five strategies where I might have developed one, and suggest dozens of actions after a meeting. It produced work faster than I could review, prioritise or implement it.

Output became abundant. Attention did not.

A prompt does not scale

A well-written prompt can produce an excellent result. For one-off work, that may be enough. Recurring leadership work needs something else.

If I return to a similar decision three months later, I do not want the quality of the process to depend on whether I remember the perfect prompt. If the work affects a colleague, investor or client, I need more than a convincing answer. I need a dependable process.

A prompt briefs the model for one task. A skill preserves how a category of work should be done. It carries the relevant context, sequence, quality standards, boundaries and approval points into every new use.

The distinction became concrete while designing a proposed AI marketing operating system for a regulated bank. The objective was not one good campaign or a library of clever prompts. The workflow needed approved organisational knowledge, defined steps, source traceability and control rules. It needed to separate confirmed information from assumptions, and to preserve human approval before anything was published.

The work was still at the proof-of-concept design stage. Its purpose was to test whether a specialist’s way of working could become a shared organisational capability.

The prompt starts the work. The skill governs how the work is done.

Knowledge makes the skill specific

Skills are useful only when they can access reliable context. Leadership depends on information accumulated over time:

• what the organisation is trying to achieve;
• what has already been decided;
• which constraints are real;
• who owns which responsibility;
• which sources can be trusted;
• what remains uncertain.

A normal AI conversation holds only a temporary, incomplete version of that. I noticed this quickly because I was applying AI across founder work, advisory engagements, investments, property operations and personal decisions. Without clear boundaries, information moved between projects, and a model could answer confidently without knowing what was missing.

Context could no longer remain hidden inside conversations. It had to become infrastructure.

I moved important knowledge into a structured, file-based system. Each project received its own background, decisions, tasks, risks and operating records. The files were not merely documentation. They became load-bearing parts of the operating model.

By October 2026: more than 80 registered skills across 18 operating categories.

The number is not the objective. It is evidence of how the system grows. A new skill is created when a type of work repeats often enough to justify a stable method. The possible applications are almost unlimited. My time is not.

General intelligence provides breadth. Private knowledge provides specificity. A skill provides method. Human judgment provides accountability.

Learning becomes part of the system

The difference between a prompt and a skill becomes most visible after something goes wrong. A corrected prompt disappears with the conversation. A corrected skill does not.

When a new fact is discovered, it goes into the knowledge base. When the method improves, the skill is updated. The next use begins from the improved version. Not every correction should become a permanent rule; the question is whether the learning is local or reusable. If it is reusable, it should not remain hidden in a chat.

The same principle extends beyond one person. The opportunity for an organisation is not to give every employee a generic assistant. It is to turn the repeatable parts of expert work into shared skills grounded in the organisation’s own knowledge, so that experts spend less time repeating routine work and more time on difficult cases and decisions.

Parallel work concentrates leadership

Once context and skills became dependable, more work could run in parallel. Separate projects move at the same time because their contexts and methods stay distinct, each producing work that can be reviewed against the same standards.

Every workstream eventually returns to a human decision point. Does the analysis reflect the real situation? Is the evidence sufficient? What is still missing? What should stop?

The queue of work waiting for judgment can grow faster than the queue waiting for production. Parallelism without prioritisation simply produces a larger inventory of unfinished thinking.

Speed also makes boundaries more important. A weak assumption can travel from research into a presentation, and from there into an apparent company commitment. The system must distinguish analysis from decision, draft from approved communication, proposal from commitment. AI can prepare, analyse and challenge. It should not silently approve its own work.

Human approval is not an inconvenience added after the system is built. It is part of the architecture.

Where leadership begins

Productive laziness was the starting motivation: attention should not be spent rebuilding work that can be systematised. The 3+ philosophy still provides the execution discipline. The Leadership OS connects them, and it remains under construction because the work keeps changing.

AI will continue to lower the cost of producing work. It cannot decide what deserves attention, or who is accountable for what happens next.

Which brings me back to those mornings.

The machines are rarely the bottleneck. I am. That is not a failure of the system. It is where leadership begins.

In brief

• AI makes production abundant; informed attention remains scarce.
• Skills and private knowledge make recurring work more dependable.
• Human judgment remains responsible for direction, approval and accountability.