The Prompt Engineering Hangover: Stop Micromanaging Your LLM

There is a weird hangover from the first wave of AI.

We were all told that the power was in the prompt.

Use this structure.

Add this role.

Say “act as”.

Give it examples.

Tell it to think step by step.

Threaten it with its life (okay maybe that one was just me).

A lot of that made sense at the time.

The models needed more hand-holding.

But that is starting to change.

With newer models like GPT-5.5, Claude Opus 4.7, and the general move towards more agentic AI, the skill is moving up a level.

You do not need to micro-manage as much.

You need to give the AI an objective or a goal.

And better context.

We need to be asking ourselves:

“What does AI need to understand before it starts?”

Because there is a big difference between giving AI a task and giving AI a proper brief.

A task is:

“Improve this product page.”

A brief is:

“This product has high margin, but customers worry about fit. Returns are expensive. We already tested urgency and it made the page feel cheap. The main drop-off is before Add to Cart. Customer support keeps getting questions about sizing and delivery. We need to increase confidence without making the brand sound pushy.”

Those two inputs will not get the same output.

This is the same as hiring someone into your business.

You would not bring in a new marketing manager and say:

“Improve conversions.”

Then leave them to it.

• You would explain the business.

• You would show them what customers care about.

• You would tell them where the money is made.

• You would explain what has already been tested.

• You would show them what the brand should sound like.

• You would tell them what success looks like.

Without that, they have to guess or worse assume.

And we all know that assumptions are the mother of all mistakes.

AI is no different.

Most AI slop is because of a lack of context.

That is why the all the AI big wigs are moving towards this concept of memory.

Not memory in the “it remembers your favourite sandwich” way.

Useful memory.

The stuff that stops it making the same mistake twice.

Some agent tools are already doing this.

OpenClaw has a feature called Dreaming.

After the agent has done some work, it looks back at what happened and asks:

• What worked?

• What went wrong?

• What new context was added?

• What should I remember next time?

Then it saves it into memory.

So the next task does not start from zero and it constantly self improves.

Andrej Karpathy has talked about a similar idea with LLMs and Obsidian.

Using AI to build a living wiki.

A place where knowledge builds up over time.

AI gets more useful when it has something to work from.

This stuff is super complex and techy right now

But you do not need to start with agents, wikis, just start with the boring version.

A simple Business Context .md file.

That is what I shared yesterday.

It gives AI the basics before you ask it to do anything useful.

So when you ask it to review a product page, write landing page copy, shape an email flow, or find CRO opportunities, it is not guessing from a standing start.

You can grab the Business Context .md file here 👉 Business Context .md File

Chat soon,

Peter