Fundamentals

Context Engineering, Not Prompt Engineering

The skill that outlasts every model release: getting the right information in front of the model

Everyone argues about the top two layers. The bottom one decides what you get back.

The Idea

Every few months a new model drops and everyone piles onto it. Fair enough — they genuinely are better. But the single most important thing you can do with AI has nothing to do with which model you're on. It's how you communicate with it, and that skill carries over to every model you'll ever use.

For a couple of years the name for that skill was prompt engineering: the wording, the structure, the order you put things in. Some of it still has an effect — knowing when "you are an expert" helps and when it hurts, or showing the model an example of what you want the output to look like. But the models have got smart enough that the wording isn't the bottleneck anymore.

cleverer wordingmagic phrasesthe right context at the right time

What kills your output is the information you never gave it. The people building these models have a name for that layer, and it's the one they care about more than your phrasing: context engineering — making sure the model has the right context at the right time. The model can only work with what's in front of it. Hand the smartest model on earth three lines and a vague goal, and it fills every gap with a guess.

The Two Questions

Before every prompt that matters, ask yourself these two. They're the whole practice in miniature.

1

What would a human need to know to do this task?

If you handed this job to a smart person who'd just walked in, what would you have to tell them? The goal, the constraints, what's already been tried, what good looks like when it's done. The model needs exactly the same briefing — it just never complains when you skip it. It answers anyway, worse.

2

What do I know that the model doesn't?

You carry context you don't even notice you have: the client hates jargon, the deadline moved twice, the last version broke in a specific way, you've secretly already decided the answer. None of that exists for the model until you say it. The gap between your head and the prompt is where most bad outputs come from.

Make the Model Do the Asking

The honest problem with "give it more context" is that you don't know what you're forgetting to say. So flip it: make the model interview you. Paste this at the start of any task that matters, exactly as it is — the model asks, you answer, and the context gets pulled out of your head instead of relying on you to remember it.

The interview prompt — paste as is
Before you start this task, interview me about it, one question at a time, until you properly understand three things: what I am actually trying to achieve, what a person picking this up cold would need to know, and anything I know that you don't yet, like constraints, history, preferences, or things that have already been tried and failed. Then play back your understanding in a few lines and wait for my go ahead before doing anything.

Make It Automatic

Pasting the prompt every time works, but you'll forget. The real setup is to make it a standing rule, so Claude gathers the context itself before every proper task without you asking. Paste this and Claude writes the rule into your CLAUDE.md and shows you the line it added — you don't touch the file.

The standing rule — paste as is
Add a short standing rule to my CLAUDE.md so this happens on its own from now on. The rule: before starting any task that isn't trivial, interview me one question at a time until you understand what I'm trying to achieve, what someone picking this up cold would need to know, and anything I know that you don't yet. Then play your understanding back before you start. Put it in whichever CLAUDE.md fits best, keep it to a few lines, and show me the exact lines you added.

Now every session starts by pulling the context out of your head. That is context engineering running on autopilot.

The Pro Tip: Dictate

Here's why nobody actually does any of this: typing is a really slow medium for getting information into a model. You know all the context — but typing three paragraphs of background before every task is a pain, so you send the one-liner and the model pays the price.

So talk instead. Dictate the prompt: what you're working on, what's gone wrong, what you actually want back. A minute of talking carries more context than most people type in a day. And don't worry if it comes out messy — these models are genuinely good at understanding your reasoning, so a rambling chain of thought isn't noise, it's useful context about how you're thinking.

I built a dictation tool for exactly this, and you can have it: Dictate — completely free, open source, works in any app on your Mac. One paste into Claude Code installs it for you.

The Honest Bit

Context engineering won't rescue a task you can't describe. If you don't know what done looks like yet, more context is just a longer version of not knowing — sometimes the right first prompt is "help me work out what I actually want here."

And it's the right information, not all of it. The window the model reads is finite, and stuffing it with everything you have works against you — that's its own topic, covered in the context window guide below. Wording still matters at the margins too, for tone and format. It's just the smaller half.

Go Deeper