What is context engineering? The prompting shift to start using now
The prompting guidance changed. Did you notice?
In the last few months, Anthropic, OpenAI, and Google have all updated their guidance on how to get better output from their AI tools.
The direction is the same across all three: the way most people have been writing prompts is now working against them.
What the old guidance said
For the past two years, the received wisdom was clear. Write longer prompts. Specify every step. Tell the AI exactly how to think. Add rules for every edge case. Be precise.
That approach made sense when the models needed more hand-holding. It doesn't any more.
What the new guidance says
Anthropic's updated documentation puts it plainly: assume the model is already smart. Only give it what it cannot infer.
OpenAI says shorter, outcome-first prompts outperform process-heavy ones. Over-specifying the steps adds noise and produces answers that are mechanically correct but miss the point.
Google goes further. Stop describing tone with adjectives, they say. Upload an example of your past work instead and let the model infer the style automatically.
The shift has a name. It's moved from prompt engineering to context engineering. Less focus on how you phrase the request. More focus on what context you give before you make it.
What this means in practice
Most people's prompts have the wrong problem. They're too long in the wrong places and too thin in the right ones.
They describe how Claude should think rather than what they need. They use adjectives — "professional but approachable, punchy but not too informal" — when one paragraph of their own writing would do more work. They treat the first output as finished rather than diagnosing what went wrong.
Three things make the difference.
Context first
What does Claude not already know? Your industry, your audience, your constraints. Not what B2B means. The information it genuinely cannot infer without you.
Task second
What do you need? Be specific about the output, not the process. "A 200-word LinkedIn post in first person, ending with a question, no hashtags" is a task. "Write something engaging for LinkedIn" is a wish.
An example, not a description
Claude is exceptionally good at matching style. One paragraph of writing you like — your own, a piece you admire — does more work than any number of tone adjectives.
The free skill
I've built a Claude skill that applies this thinking every time you ask it to write, review, or improve a prompt.
It asks the right questions before building anything. It tells you what to upload rather than what to type. And when the output goes wrong, it gives you a diagnostic — so you change one variable rather than rewriting everything from scratch.
The skill is free, access it here. If it doesn't prompt you to install it when you download it, open Claude and go to settings. Select customise, then skills. Click +, choose create skill, upload and save.
If you'd like help building prompts, system instructions, or AI workflows for your business, get in touch.