Two years ago, “prompt engineer” was showing up on job descriptions. Today, the models write better prompts than most humans do. That transition happened faster than anyone predicted, and the implications are worth thinking through.

What died

Prompt engineering, in its original form, was about coaxing a model into useful behavior through clever phrasing. Specific trigger words. Magic prefixes. Chain-of-thought incantations. It was a workaround for models that were too rigid to interpret intent.

Modern frontier models don’t need that. They infer intent well. They handle ambiguity. They respond to plain language from non-technical users better than they respond to baroque prompt structures from self-declared prompt engineers. The craft, as a distinct skill, is being commoditized.

More bluntly: if you give a model a rough description of what you want, it will often write a better prompt for itself than you wrote for it. Meta-prompting is now a standard technique. The model is better at this than you are.

What actually matters now

Context engineering is what separates good AI outcomes from mediocre ones. Not the phrasing of your request, but the information environment you build around it.

This means:

  • What does the model know about the task before you start?
  • What does it know about the system it’s operating in?
  • What constraints, examples, and prior outputs are in scope?
  • What does it not need to know, and are you keeping that out?

The quality ceiling for AI output is largely set by context quality. A well-phrased prompt with poor context will consistently underperform a plainly worded request with rich, relevant context. This is not a marginal difference — it is often the difference between useful and useless.

Skills as Standard Operating Procedures

One pattern I keep coming back to: the best way to think about reusable AI tasks is as Standard Operating Procedures.

In traditional organisations, an SOP captures how a task is done — the inputs, the steps, the quality criteria, the expected output format. It removes dependence on individual knowledge and makes outcomes repeatable.

A well-written skill for an AI agent is exactly this. It defines:

  • What the task is
  • What context the model needs
  • What constraints apply
  • What the output should look like

Done well, a skill means anyone — or any agent — can invoke a complex, nuanced task reliably, without needing to know how to prompt it from scratch. The skill is the institutional knowledge, encoded for AI execution.

This reframe matters for organisations building AI capability. Stop thinking about prompts as one-off interactions. Start thinking about which tasks are worth encoding as reusable, auditable procedures — and what it takes to make those procedures reliable enough to trust.

This is a field note from my DBA research on AI enterprise architecture. If it resonates, subscribe for more.