Prompt, loop and graph engineering: is it all just jargon?

1Magic prompts

What actually happens when Riya types a question? Her words go straight into the context, and the word-guesser carries on from there. Change the note, and you change the guesses. So people started asking: which note gets the best guesses?

Like kids hunting for cheat codes in a video game, people everywhere tried strange phrases. "You are a world-class expert." "I'll tip you $200 for a perfect answer." "Take a deep breath." Some of them seemed to work. In 2023, a computer search at Google found that "Take a deep breath and work on this problem step-by-step" beat "Let's think step by step" on school maths for one of its models: about 80 right out of 100, against 72. In 2024, another search found that one model did best at maths when its instructions read like a Star Trek mission, starting with "Captain's Log". Seriously.

A prompt is simply the note you give the AI. That hunt for the perfect wording is what most people still mean by prompt engineering.

But why would "think step by step" help a word-guesser at all? Try this one yourself: a bat and a ball cost ₹110 together. The bat costs ₹100 more than the ball. How much is the ball?

Your gut says ₹10, doesn't it? Most people's does. When psychologists gave students the dollar version of this puzzle, more than half of those at Harvard, MIT and Princeton went with the gut answer. It's wrong: if the ball cost ₹10, the bat would cost ₹110, and together they'd cost ₹120. The ball costs ₹5.

A model that jumps straight to the answer has the same problem. It's guessing the whole result in one go, and the gut answer is the loudest one. But every word it writes lands in its context and becomes part of what it reads for the next guess. So if it writes its working first (the ball is x, the bat is x + 100, together that's 110, so x is 5), each small step is right there to lean on, like the rough work in the margin of Riya's maths paper.

2Tricks that expire

So why did the magic stop working? Because the people who build AI models noticed too. If writing the working helps, why wait for someone to ask? So in September 2024, OpenAI released a model called o1, trained to think before it answers: it does its rough work privately, then replies. On a tough American maths competition called AIME, OpenAI's previous model, GPT-4o, solved about 12% of the problems. o1 solved 74% on its first try. Today, most new AI models can do this on their own. The five words moved inside the model.

And once that happened, the five words stopped mattering. OpenAI's own guide for these thinking models says asking them to think step by step is unnecessary. In 2025, researchers at Wharton tested it: for thinking models, the request moved scores by only about three points, up or down, and made answers 20 to 80 percent slower.

What about the other tricks? Researchers tested those too.

1. Tips and threats. In 2025, the same Wharton team offered AI models a $1,000 tip, and even a trillion-dollar one. Overall: no real difference.
2. "You are an expert." A 2024 study tried 162 different roles, from experts to friends, on 2,410 fact questions. Adding a role didn't make the answers more correct.
3. SHOUTING. People used to write "CRITICAL: You MUST…" so older models would listen. Anthropic's guide now warns that its newer models listen so well that shouting makes them overdo it. Just say it normally.
4. The Star Trek prompt. It was the winner for one model. Other models had completely different winners.

Is this just bad luck for prompt fans? No. It's one of the oldest patterns in AI. In 2019, the AI researcher Rich Sutton wrote a short essay called The Bitter Lesson. Looking back over 70 years of AI research, he saw the same story again and again: clever human shortcuts win for a while, but in the long run, general methods that use more computing power win, by a large margin. Chess programs built on human chess wisdom lost to programs that simply searched deeper, and in 1997 one of those beat the world champion.

Magic prompts are exactly that kind of shortcut. Every trick that reliably makes the model think better gets trained into the next model. So a list of secret prompts is like last year's answer key: by the time you buy it, the paper has changed.

3What goes in the context

But if tricks expire, what doesn't? The things no update can train into the model, because they only live in your head. Let's go back to the context. Whatever the model does, it can only use what's in there. It can't see Riya's notebook, her teacher's instructions or her last essay, unless someone puts them in.

Say Riya needs an essay: "My cricket hero", 300 words, due Monday. What should the app put in the context? Her teacher's instructions: yes. Her last essay, where the teacher wrote "Loved your personal story!": yes, because it shows what good looks like. Last week's chat about pizza toppings: no. Fifty pages of cricket stats: no.

Wait, why not just put everything in? Because the context has a size limit, and even before it's full, clutter hurts. Researchers have found that models often miss facts buried in the middle of a long context, and that answers get shakier as the context grows. A tidy context beats a full one.

Choosing what goes into the context, and what stays out, is called context engineering. It's the grown-up version of prompt engineering. By 2025, people building AI apps had started using that name, because they'd noticed that the prompt is just one thing in the context.

4Look it up first

Here's a question the word-guesser can't answer. Riya asks her school's AI helper: "When's our championship match?" The model learned from text written months ago, and it has never seen her school's notice board. So on its own, it would guess. And a guess can sound very confident.

So what does a smart student do in an open-book exam? She doesn't guess: she looks it up. The app does the same. Before the model answers, the app searches the school notices, finds the right one and adds it to the context. Now the model simply reads it out, and it can even say where the answer came from.

The clever part is how it finds the notice. The notice doesn't say "championship". It says "inter-school cricket final". So the app doesn't match words. It matches meanings. Every piece of text gets turned into a point on a kind of map, where things that mean similar things sit close together. Engineers call those points embeddings. "Championship match" lands right next to "inter-school final", and far away from "pizza order".

That's all RAG is. The long name, retrieval-augmented generation, comes from a 2020 research paper, and it's just our three steps: retrieval (look it up), augmented (add it to the context) and generation (write the answer). It's a mouthful for "open-book exam".

Notice something? RAG is really one kind of context engineering: putting the right page into the context at the right moment.

5Try, check, try again

How does Riya get better at bowling? She bowls one in the nets, sees it drift down the leg side, adjusts her wrist and bowls again. Bowl, look, adjust. Nobody gets it right with the first ball.

AI works better the same way. Remember the order slips? The model can order a check: run the code it wrote, count the words, search for a fact. The result lands in the context, the model reads it, fixes what's wrong and tries again, until the check passes.

Riya asks for a five-question cricket quiz with one-word answers. The model writes it, then orders a word count. Two answers are too long! So it fixes them, checks again, and this time everything passes.

That cycle is called the loop. A word-guesser working in a loop, with a goal and some tools, is what most people mean by an AI agent, and that's a story big enough for a post of its own. Deciding what the loop checks, when it stops and when it must ask a human first got its own buzzword in 2026: loop engineering.

Companies do the same checking at scale. Before trusting an AI with real customers, they write a test: a hundred tricky cases with known right answers. Then they run every new version through it, and see what breaks. Those tests are called evals.

6The route map

But what if you don't trust the AI to choose its own steps? Then you draw the route for it. Picture a metro map. Station one: read the question. Station two: search the notices. Station three: write the answer. Station four: check it, and if it fails, go back to station three. The train can only go where the tracks are laid.

In tech-speak, the stations are nodes, the tracks are edges, and a map of nodes and edges is a graph. So drawing these maps for AI apps is called graph engineering. Builders also call them workflows.

Sometimes a fixed route is exactly right. A bank wants every loan form to pass the same checks, in the same order, every single time. But a train can't leave its tracks. Ask something the map never planned for, like "Did our team score more runs than last year?", and there's no station for the scorebook or the calculator. The train runs its usual route and gives a so-so answer.

Smarter models can work more like a taxi driver who knows the city. Give them the destination and the tools, and they pick the route. One engineer who builds AI research tools shared a telling story in 2025. He first built his research helper as a fixed flowchart in late 2024, because that was what the models of the time could handle reliably. As the models improved, the flowchart became the very thing holding his helper back, so he rebuilt it to let the model plan. Anthropic's guide for its own models says something similar: a simple "think thoroughly" often works better than a step-by-step plan written by hand.

7What never expires

So is all of this useless? Not at all. But look closely at every trick that expired. Each one tried to make the model think harder or try harder, and that's exactly what the model makers keep improving. There's a short list of things no update can ever add, because they live in your head:

• What you want: the goal, said clearly.
• What you know: the context only you have.
• What good looks like: an example.
• Whether it's right: checking the result.

That's the whole secret, and it isn't a secret at all. The AI can learn every trick. It can't learn what's in your head.

Riya already knows this from cricket. Before the last over, her coach doesn't whisper a magic prompt to the bowler. He says: "We need him caught at slip. The pitch is dry, and he chases anything outside off stump. Bowl it like the one that got him last match." Then he watches every ball and moves the field. Goal, context, example, check. No cheat codes.

So Riya tries it herself. She has to give a one-minute speech to her team before the final.

That's why the AI makers' own guides sound like advice for managers. Anthropic tells people to treat its model like a brilliant new employee who doesn't yet know how your workplace runs, and to test a prompt by showing it to a colleague: if they'd be confused, the AI will be too. Google's beginner course on prompting boils down to: say the task, give context, add references, check the result and try again. Good managers have been doing context engineering for decades. They just called it giving a clear brief.

And the AI is learning to meet us halfway. Some AI tools now ask you a few questions before they start a big task, to fill the gaps in your note.

8So what does this mean for you?

Let's zoom out. Here are the five buzzwords, one picture each:

• Prompt engineering: hunting for magic prompts. The models keep absorbing them.
• Context engineering: choosing what goes into the context, and what stays out.
• RAG: looking things up by meaning, with embeddings, and adding them to the context.
• Loop engineering: try, check, fix, and know when to stop. Evals are those checks, at scale.
• Graph engineering: a route map drawn in advance.

Are the people who use these words fooling us? Mostly, no. Builders need names for their work, the way cricketers say "yorker" and "doosra". Those words are handy in the dressing room. But you don't need them to enjoy a match, or even to bowl a great ball.

And there'll always be a new one. Prompt engineering had its big moment around 2023. Context engineering arrived in 2025. Loop and graph engineering are 2026's. But the models keep getting smarter: they think for as long as they need, look things up, check their own work and pick their own route. So the jargon will keep changing, but the skill underneath won't.

So the next time a reel promises secret prompts, ask one question: is this a trick, or a habit? Tricks expire. Habits don't.

So, those five words from 2022. They worked because a word-guesser does better when it can read its own rough work. They stopped mattering because the models learned to do the rough work themselves. Every magic prompt that works gets trained into the next model, so the tricks keep disappearing, and they aren't coming back. What you want, what you know and what good looks like can't be trained in, because they're about you, not the model. That's why prompt engineering lost to context engineering, for good.

As for Riya, she skipped the ₹1,999 course and bought new batting gloves.

Her uncle, Ravi Mama, watched the quiz fix itself over her shoulder, and it gave him an idea. He edits wedding videos. He has 212 clips, and a client who wants a highlight reel by breakfast. If the AI can check and fix its own work once, why not let it keep going, all night, with nobody watching?