Two years ago, the difference between a useless AI answer and a brilliant one was usually the prompt. Today's models are good enough that clever phrasing barely moves the needle — you can ask politely or bluntly and get roughly the same quality. What actually separates people getting real work out of AI from people getting generic mush is something less glamorous: what information they put in front of the model before they ask. That's context engineering, and it's the skill worth learning now.
Why Prompt Tricks Stopped Working
The old prompt-engineering playbook was full of workarounds for weak models. "Take a deep breath." "You are a world-class expert." "Think step by step." Those existed because early models needed nudging into a better mode. Modern models reason by default, so the nudges are mostly noise. Meanwhile the real failure mode has stayed exactly the same and got more visible: the model gives you a confident, well-written answer built on nothing, because you never told it anything specific about your business, your customers, your data, or your constraints. Better wording can't fix missing information.
The Four Things Every Serious Request Needs
When I set up an AI task now — writing listings, drafting a lesson, analysing a spreadsheet — I check four boxes before I care about phrasing. Source material: the actual documents, exports, transcripts or examples the answer should be built from. Standards: what "good" looks like, ideally three examples of past work you were happy with. Constraints: word counts, banned phrases, policies, tone, things it must never claim. Task: the one specific thing you want back, in a stated format. Miss any of the first three and you're asking a stranger to guess your business.
Curate Ruthlessly — More Context Isn't Better Context
The instinct once you understand this is to dump everything in. Don't. Long context windows have made it possible to paste a hundred pages, but relevance still beats volume: bury three important lines inside forty pages of filler and the model will weight the filler. Before pasting, ask yourself what a smart freelancer would actually need to do this job on day one. That's usually two or three tight documents, not your entire Drive. If you're pasting something long, tell the model what matters in it and what to ignore.
Build Reusable Context Packs
This is where it stops being a chore. Most of your context is the same every time, so write it once. I keep a small set of plain-text files — a brand voice file with real writing samples, a policy file with returns and delivery terms, an audience file describing who I'm actually talking to, and a spec file per business. Every new task starts by attaching the two or three relevant packs. Setup drops from ten minutes to ten seconds, and quality becomes consistent instead of depending on how much effort I put in that particular morning.
Give It Somewhere to Say "I Don't Know"
Hallucination is usually a context problem wearing a costume. A model invents a delivery policy because you asked about delivery and gave it no policy. The fix is two lines in your setup: state that the attached documents are the only valid source for facts, and give it an explicit escape route — say "not in the provided material" rather than guessing. It feels almost too simple, but it converts a confidently wrong answer into a flagged gap, which is a completely different problem to have.
How to Practise This Week
Take the AI task you do most often and do it twice. First the way you normally would, from a blank chat. Then again with a proper context pack — source material, three good examples, your constraints, the task stated once and precisely. Compare the two outputs and count how many edits each needs. In my experience the gap is not subtle, and once you've seen it once you stop opening blank chats forever. Prompting was a trick. Context is a system — and systems are what actually scale.