Why "prompt engineering" is a terrible name.
It is not engineering. It is not even prompting. The name sends people to the wrong place to fix a problem, and that costs them money and time.
Here is what the work really is, in plain words, and why the label matters more than it should.
Good wording is a small part. The other parts decide if the answer is useful.
- Not engineering. Engineering means a spec, a test, and a result you can repeat. Most "prompt engineering" has none of the three.
- Not even prompting. A prompt is the last thing you type. The real work happens before you type it and after the answer comes back.
- What it is. Three jobs: know what you actually want, set the limits, and check the answer against both.
- Why the name costs you. If wording is the problem, you hunt for better wording. If the problem is intent, you can rewrite forever and still miss.
It is not engineering.
Think about what a bridge engineer does. They write down what the bridge must carry. They test it. If it fails, they find out why and fix the cause. The same bridge comes out the same way next time.
Now think about the usual prompt advice. "Add the words 'think step by step'." "Tell it to act as an expert." "Use this template from a video." Nobody wrote down what the answer must do. Nobody tested it on ten cases. If it works, nobody knows why. That is a lucky guess with a good vibe.
There is a second problem. Models change. A phrase that worked last quarter can do nothing this quarter. A craft built on exact phrases breaks every time the engine under it changes. Engineers do not build on sand.
It is not even prompting.
The prompt is the part you can see. It is the smallest part of the job.
We have a test for this on the site. We call it the Don't Kick Test. A child hears "don't hit your brother" and kicks him instead. Then the child says, "You did not say not to kick." The child followed the words. The child missed the point.
AI does this all day. It follows your words and misses your intent. You can fix the words, and the next answer finds a new kick. The words were never the lever. What you wanted, and what you would never accept, were missing from the request.
"Write a short reply to this unhappy customer."
Keep the customer. Do not promise a refund. Do not blame the shipper. Sound like a person, not a form letter.
Better wording will not carry the second box. Only you hold it. The job is getting it out of your head and into the request, and then checking that the answer honors it.
What it actually is.
Three jobs. None of them is about clever phrasing.
Find the real intent
The literal request and the real goal are often different. Ask what you want the answer to do for you. Our own interview material runs on this idea: structured questions that dig out who you are and what you want, so the model stops guessing.
See the five filesSet the limits
Say what is out of bounds, how long the answer should be, and who it is for. Limits turn "good" into something you can check. Short, specific questions beat one long paragraph of adjectives.
See the vaultCheck the answer
Compare what came back to the intent and the limits. Do not ask the model if it is right. A model told to confirm will usually confirm. Verify against something outside it.
See the methodA better name for the whole thing is intent and checking. It is less catchy. It points at the right work.
A wrong name sends you to the wrong fix.
You buy tutorials that do not fit your job.
If wording is the skill, then someone else's template should work for you. It rarely does. Their context is not your context. Our homepage lists "paste prompts from tutorials that do not fit your context" as the old way, and for good reason.
You edit the same output again and again.
When the intent is missing, every answer needs heavy editing. You pay for that in hours. The fix is upstream, in the request, and no phrase tweak reaches it.
You chase perfect wording that does not exist.
There is no magic sentence. If your tool needs one exact phrasing to behave, the tool is fragile. A good setup works when you talk in plain English, and that is how we describe ours: your words go in, no syntax, no learning curve.
You trust an answer that only looks right.
Words can be perfect and the answer can still be wrong. Our method page names one small cause: a prompt that asks the model to confirm rather than verify. Naming the job "prompting" hides that checking is part of it.
You lock your know-how to one vendor.
Phrases tuned to one model go stale when you switch. Intent, limits and checks are yours. They can move with you. That is the idea behind the portable doctrine on the method page.
The work, step by step.
This is what we do instead of hunting for better phrasing. Each item below exists on the site today.
Intent interview
A structured interview, twenty questions plus writing samples, builds a short profile of who you are and how you write. It loads at the top of a session. Details on the five files page.
Intent mapping
A check that digs for the real intent behind the literal request. It is one of the five checks in the vault.
Pre-send cleanup
A pass that runs before every send. It rewrites messy requests with action verbs and length limits, and flags an answer that is heavier than the question deserved.
The Don't Kick Test
Paste a prompt and the answer. The page shows where literal compliance hides a miss. It makes no LLM call, and it is free to run. Try it.
The one-line version.
result = f(intent, constraints, verification) and
wording is a weak term in that function.
Hold intent and limits fixed, and many different phrasings land close to the same place. Change the intent, and no phrasing saves you. Add a check outside the model, and a bad answer gets caught before it ships. That is the whole argument. It is also why we think the field needs a better name.
Where to go next
The method for how we check work, and the questions page for plain answers to the usual worries.
Stop polishing words. Fix the request.
Bring us one prompt that keeps missing. We will show you what the real intent was, and what check would have caught the miss.