Beyond Prompt Engineering

Prompt engineering had a short golden age. Around 2023, the gap between a useless AI answer and a good one really could come down to wording. Telling a model it was a world-class expert actually helped. So did lines like 'think step by step', made-up roles, and long templates that people tuned for months and guarded like secrets. Job ads offered six figures for it.

Most of those tricks do nothing now. Not because people got worse at writing prompts, but because the thing on the other end changed. A request that used to look like this:

You are a world-class senior software engineer with 20 years of experience. Take a deep breath and think step by step. It is very important to my career that you get this right. Now, carefully review the following code for bugs...

Is now just: 'check this for bugs.' Same answer, often better. Everything in that first version was a workaround for a system that could not yet handle normal speech.

Why the Tricks Stopped Working

Three things happened around the same time.

1. Models got better at messy questions. Early models were fragile. Small changes in wording swung the answer a lot, which is exactly why hunting for the perfect phrase paid off. Today's models are trained hard on following plain instructions, so a vague, half-finished request works about as well as a polished one.
2. The chat box stopped being the only way in. You can now hand a model a schema and get back data in that exact shape, instead of asking nicely for clean JSON and hoping. You can give it tools it calls directly, so a request turns into a series of real actions rather than a wall of text. And plenty of input is not text at all now - images, voice, screens.
3. Models write better prompts than we do. If you let a model draft its own instructions and score them against real examples, it beats people doing the same thing by hand. We were slowly turning a dial the machine could turn itself.

The skill didn't die because it got easy. It died because it got automated, and because the hard part moved somewhere else.

Context Engineering

The hard part moved from how you ask to what the model can see when you ask. People call this context engineering. It is a data problem, not a writing problem.

Most AI tools now look things up before answering - they search your documents, pull back the relevant bits, and hand those to the model along with your question. When the answer is wrong, the lookup is usually what failed, not the wording. The model read the wrong page and answered it correctly. So the questions that decide quality are dull ones:

• Is it pulling the right documents for this question?
• Are they current, or a copy from eight months ago?
• Is there enough to answer - and not so much that the useful part gets buried?
• Can anything in there be trusted?

That last one matters more than it sounds. Anything you pull in - a web page, a shared doc, a support ticket, an email - might contain text written to trick the model. If your system treats what it read as orders rather than as information, someone can plant those orders. No amount of clever wording in your own prompt fixes that.

Loop Engineering

The second change is about shape. The old idea was one question in, one answer out, and the answer had to be right immediately. That is a bad bet, so the tools that work stopped making it.

What works instead is a loop. The model tries something, checks the result against something real, and goes again. With code this is easy to picture: write it, run it, read the error, fix it, run it again. The tests pass or they don't. The computer does not care how confident the answer sounded.

A single answer has to be right immediately. A loop only has to tell right from wrong, and be allowed to try again.

That check is the whole trick, and it explains why AI agents feel amazing at some jobs and useless at others. If there is a cheap, honest way to tell whether an attempt worked, the loop gets closer to right every time around. If there isn't, it just goes in circles.

• Works well: code that runs or crashes, tests that pass or fail, a database query that returns rows or errors, a scraper that either found the price or didn't.
• Works badly: is this ad copy good, is this summary fair, does this design look right. Nothing can score the attempt, so nothing improves.

If you want to guess whether an AI system will work before you build it, ask what plays the role of the tests. If nothing does, you are back to one-shot answers with extra steps.

The Boring Plumbing

Once software takes actions instead of just writing text, the hard questions become normal engineering questions:

• What is it allowed to touch?
• What passwords and keys does it hold, and are they limited to only what it needs?
• What happens when it gets something wrong - can you undo it, and does a person check first?
• What gets written down, so you can work out later what went wrong?

None of that is prompting. It is the same care you would take with any system that has real permissions. The one difference is that this one does not behave the same way twice, which makes the guardrails matter more, not less.

The Part That Didn't Change

This pattern is not new. People once memorised search tricks - quotes, minus signs, site filters. Then search got good enough that you could type what you wanted. The trick knowledge disappeared into the tool, and the people who had it moved up a level.

What survives every time is knowing what you actually want, clearly enough to say it, and being able to tell whether what came back is right. That second half keeps getting harder, because AI writing is now smooth enough that wrong answers read exactly like correct ones. Being able to judge an answer is worth more than being able to coax one.

So the short version:

• Don't collect prompt templates. That knowledge expires in months.
• Learn how your tools decide what to look up, because that is where most bad answers come from.
• Build loops that can check their own work, and be honest when a job has no way to check.
• Get good at spotting a confident wrong answer. No software update is going to take that job from you.