My AI assisted me in writing this article.
But I did not write a carefully structured prompt. That's why I believe something about how we work with AI has changed.
For the past few years, much of the conversation was about prompt engineering. People created prompt formulas, prompt templates, role-playing instructions, and increasingly complicated ways of telling an AI what to do.
Now context engineering is the new hot phrase, but, Context was always part of the prompt.
OpenAI defines prompt engineering as designing and optimizing input prompts to guide a language model's responses.
So, what's different?
You might also like: Prompt Engineering Guide for Nigerians
What Is Context Engineering?
Context engineering is the deliberate process of giving an AI the information, constraints and surrounding state it needs to produce the desired result.
That information can include the current instruction, previous conversations, examples, documents, memories, retrieved information, tool results, system instructions, and other data available to the model when it responds.
Anthropic describes context engineering as the natural progression of prompt engineering: prompt engineering focuses mainly on writing and organizing instructions, while context engineering deals with curating all of the information available to the model during inference.
Context engineering has become broader than prompt engineering, especially when we move from simple conversations with AI into AI systems and agents.
So, Is Context Engineering Really Different From Prompt Engineering? My answer is YES.
Why Context Engineering Is Getting Attention
The old obsession with perfectly structured prompts made more sense when models needed more hand-holding.
Modern models are far better at understanding ordinary language, long instructions, and conversations that develop over multiple turns.
The better the model becomes at understanding natural language, the less important prompting becomes and the more important the information itself becomes.
My 3 Step Context Engineering Framework For LLMs
1. Describe the Problem
Start with the actual problem and not a fake role like "act as the world's greatest strategist."
Tell the model what is happening and what needs to change so the model knows what problem it is actually solving.
There is an important difference between describing a task and describing a problem.
"Rewrite my website homepage" is a task.
"People visit my homepage but cannot immediately understand what we do" is a problem.
The second gives the model more room to reason about the desired result.
2. Give It the Right Context
Now supply the information that would help someone solve that problem.
Who is the customer?
What have you already tried?
What data do you have?
What should not change?
What examples represent what you want?
What restrictions matter, etc.?
Relevant context gives AI superpowers.
3. Define the Output
Finally, tell the model what success looks like.
State what must be included.
State what should be removed.
If you already know what good work should look like, tell the model.
The entire framework is
Problem → Context → Desired Output
Experts In Any Field Can Get So Much Out of LLMs
An expert usually knows what good work looks like.
A marketer knows when positioning is weak.
A programmer understands the expected behaviour of a system.
A lawyer knows which facts are material to an argument.
A logistics operator understands what information determines whether an order can actually be fulfilled.
Knowledge allows the expert to give the AI better context and define a better destination.
Someone with very little knowledge of a subject can ask an AI to "give me the best strategy."
An expert can describe the situation, identify the variables that matter, provide the relevant data and explain exactly what the final strategy should accomplish.
The difference in output can be enormous; the advantage is not knowing a secret prompt but knowing what information matters.
Warning! Context Can Push the AI in the Wrong Direction
There is one problem with everything I have said so far. Context steers the model.
Language models are overly agreeable with users. This behaviour is called sycophancy. Research from Anthropic has found models matching a user's stated beliefs instead of giving the most truthful response, and a joint Anthropic and OpenAI evaluation also observed forms of sycophancy across models from both companies.
If I spend 2,000 words explaining why I believe an idea is brilliant and then ask the model to evaluate the idea, I have already shaped the environment in which that evaluation happens.
So good context engineering sometimes means deliberately giving the model permission to disagree.
I personally find instructions such as these useful during serious research:
"Find evidence against this."
"Assume my hypothesis is wrong."
"Separate what I believe from what the evidence establishes."
Context engineering should improve reasoning, not improve the model's ability to agree with us.
Context Engineering Is Much Bigger
With AI agents operating across many steps while collecting new information as it works.
Some information needs to remain isolated until it becomes relevant to the agent.
At this level, context engineering is clearly bigger than writing prompts.
0 Comments
Write a comment