Your process is the real prompt

During the first wave of generative AI, the advantage seemed to lie in discovering the right phrase. Collections of prompts and formulas circulated, promising to turn an ordinary request into an expert answer.

How we ask matters. But in real work, the best prompt is usually the visible part of something much larger: a process the organisation has not yet expressed.

What an expert adds without saying it

When we ask an experienced person to review a proposal, they do not receive only the document. They know the customer, earlier mistakes, commercial constraints, the tone, what can be promised and who must approve an exception.

If we give the same document to AI with the instruction “review it like an expert”, almost everything that turns experience into judgement is missing. Better adjectives do not recover that context.

From prompting to context engineering

An AI work system needs to organise several layers:

• Objective: which outcome we seek and for whom.
• State: what we know about this specific case.
• Knowledge: relevant policies, documentation and decisions.
• Examples: what we consider good, insufficient or incorrect.
• Tools: what the system can consult or modify.
• Boundaries: what it cannot do and when it must stop.
• Evaluation: how the result is verified.
• Memory: which learning deserves to be kept for next time.

Context engineering means deciding which information the model needs at each moment, not accumulating everything available. More context can also introduce contradictions and noise.

If you cannot explain the process, AI will invent it

Many implementations discover that two people perform the same task using different criteria. Until then, the variation remained hidden in emails and conversations. AI does not create that ambiguity; it makes it visible.

We can respond by writing ever longer instructions or use the opportunity to agree on the process: which inputs are mandatory, which decision is made, which exceptions exist and what must be recorded.

That organisational work is more valuable than a perfect prompt because it also improves collaboration between people.

Feedback closes the cycle

A static prompt captures what we thought we knew. A system learns when it records the cases in which the result was corrected, which criterion was missing and what change should be incorporated.

Not all feedback deserves to become a universal rule. We need to distinguish an exception from a pattern. Once again, human direction decides which learning enters the system and which context should remain temporary.

The prompt as the interface to a process

The best way to understand a prompt is not as a spell, but as an interface. It connects an intent to knowledge, tools and controls. It can be brief if the system behind it is good.

When someone shares a spectacular prompt, do not ask only about its words. Ask which data, examples, tools, reviews and decisions make it work.

Sources and references

• Anthropic’s introduction to context engineering for agents.
• Primary source: Victor’s experience training professionals and designing agent workflows.