Beyond Prompting

Prompt engineering isn't dead. It was never the right name for what mattered.

Remember GPT 3.5?

ChatGPT arrived on 30 November 2022, and the model underneath it held 4,096 tokens. About six pages. A long conversation quietly forgot its own beginning. It could write, fluently, but it couldn't think: hand it a question built on a false premise and it answered the false premise, confidently, in good paragraphs. It knew nothing about you, your company, or your problem unless you typed it in, every message, from scratch. And it drifted. It apologised, added disclaimers, wandered off the format halfway down the page.

It was remarkable and it was broken in specific ways, and if you used it seriously in 2023 you remember what we all did about it. We learned to talk to it in a certain way. We called that prompt engineering.

What it actually was is a protocol: a set of rules two parties agree on when they can't just talk. Every rule in it was our side of the partner's limitation. And the interesting thing about the three years since is not that the rules stopped working. It's that each time the model gained something, the technique moved with it, climbing a rung, until the thing left at the top wasn't a technique at all.

Rung one: prompt engineering

The 2023 model could write but not reason, held very little, and knew nothing about you. So the protocol did the reasoning's job, the memory's job, and the introduction's job, inside one message.

Act as a senior tax lawyer. The role, because one professional word shifted the entire register the model drew from. Headers, bullets, XML tags: a skeleton, because it couldn't hold the shape of a long instruction on its own. Anthropic's own documentation in July 2023 said it plainly: "Claude has been finetuned to pay special attention to the structure created by XML tags." Three worked examples, because it couldn't infer a format from a description. A list of things not to do, because it drifted and a fence was the only steering we had. And everything it needed, pasted in, every time, because it remembered nothing.

None of this was superstition, and the numbers say so. In January 2022 a Google paper showed that giving a model eight examples of worked-out reasoning lifted its score on grade-school math word problems from 17.9% to 56.9%. Four months later another paper found you didn't even need the examples. Adding one sentence, Let's think step by step, took a model from 17.7% to 78.7% on arithmetic problems. One sentence. That's the gap prompt engineering lived in, and it was real.

It was also an industry. By June 2023 OpenAI had published its own guide, and the tactics read like a summary of the year: adopt a persona, use delimiters, specify the steps, provide examples, give the model time to "think". A GitHub list of copy-paste prompts, Act as a Linux Terminal, Act as a Travel Guide, had 96,000 stars by the start of 2024 and has 171,000 today. Job seekers noticed too. Searches for "prompt engineer" on Indeed went from 2 per million in January 2023 to 144 per million by April.

Every line of it mapped to something the model could not do. We weren't communicating. We were accommodating.

Then instruction following got good enough that the exact wording mattered less, and that exposed the real problem: one message, however well engineered, could not carry real work. Real work has steps.

Rung two: flow engineering

So prompting stopped being a sentence and became a sequence. One prompt drafts, a second critiques, a third revises. A checker runs after every step. If the result fails, loop and retry. Chain of thought, which began as a phrase you typed, grew into an architecture you built.

The tooling arrived fast. LangChain, an open-source library for wiring prompts into chains, went from 5,000 GitHub stars in February 2023 to 18,000 by April and raised $10 million the same month. AutoGPT, which put a model in a loop and let it set its own next task, became the top trending repository on GitHub within days of its March 2023 release and passed 100,000 stars in weeks. BabyAGI did the same trick in about a hundred lines of Python. By December, n8n, the workflow tool people were already using to glue their apps together, shipped LLM nodes, and a prompt became a box you drag into a flowchart.

The idea got its name in January 2024, in a paper from CodiumAI titled From Prompt Engineering to Flow Engineering. The number in it is the whole rung: on a competitive coding benchmark, GPT-4 solved 19% of problems with a single well-designed prompt and 44% with a multi-stage flow. Same model. The difference was the scaffolding around it.

This was the era of chains, and it worked. It was also the last time the thinking lived outside the model.

On 12 September 2024 OpenAI released o1, "designed to spend more time thinking before they respond". On a qualifying exam for the US math olympiad, GPT-4o had solved 12% of problems; o1 solved 74%, in a single attempt, unprompted. Four months later DeepSeek released R1 under an open licence, matching it. The reasoning pipeline we had built by hand became something the model does on its own, and think step by step turned into telling a fluent speaker to sound out the letters. It doesn't hurt. It doesn't help.

What that exposed was sharper. A model that reasons beautifully over the wrong information still gives you the wrong answer. The bottleneck wasn't the thinking anymore. It was what the model could see.

Rung three: context engineering

A model only knows what's in the window, and the window used to be a postcard. So the job moved out of the prompt entirely. Not what do I type, but what does the model have access to, and when.

The technique here was retrieval: keep your documents in a database, pull the relevant ones at the moment of the question, put only those in the window. The paper is from 2020, but the market arrived in a single month. In April 2023 three vector database companies raised money in three weeks: Chroma $18 million, Weaviate $50 million, Pinecone $100 million at a $750 million valuation. That's what it looks like when an entire industry decides the prompt isn't the bottleneck anymore.

The term came later, and from the top. In June 2025 Shopify's CEO wrote that he preferred "context engineering" to prompt engineering because it "describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM". Andrej Karpathy agreed a week later, calling it "the delicate art and science of filling the context window with just the right information for the next step". By September Anthropic had published an engineering guide under that name. Prompting is about the instruction. Context is about everything around it.

And then the model absorbed most of it. The postcard became a library. Claude went to 100,000 tokens in July 2023, "hundreds of pages", and 200,000 that November. Gemini reached a million in February 2024 and opened two million to every developer that June. Memory became native: ChatGPT started remembering across conversations in April 2024 and could reference every past chat by April 2025; Claude followed that September. Projects folders let you drop your documents in once and stop pasting. What was a retrieval pipeline you built in 2024 is a checkbox in 2026.

Which exposed the next thing. A model that knows everything relevant to your problem and still can't touch it. It could see the file. It couldn't run it.

Rung four: harness engineering

Until recently, tools were human-run. The model said search for this; you searched, pasted the result, waited. It wrote the code; you ran it, pasted the error, waited. The human was the runtime.

The plumbing changed first. OpenAI added function calling in June 2023, so a model could ask for a tool in a form a program could act on. Anthropic made tool use generally available in May 2024 and, that November, published the Model Context Protocol, an open standard for connecting models to data and tools. Then the models started using screens: Anthropic's computer use in October 2024, OpenAI's Operator in January 2025.

What that unlocked is the thing most people now mean by "AI". Devin, in March 2024, was the first agent to resolve real GitHub issues end to end: 14%, against a previous best of 2%. Cursor passed $500 million in annual revenue by June 2025, used by over half the Fortune 500. Claude Code went from research preview in February 2025 to general availability in May to $500 million in run-rate revenue by September, with usage up ten times in three months. By late 2025 McKinsey found 62% of organisations at least experimenting with agents.

The technique climbed with it, into the environment the model acts in: the tools it can call, a sandbox to call them in, a loop with a stop condition, memory across runs, permission boundaries. In February 2026 OpenAI described building roughly a million lines of code across 1,500 pull requests over five months with no hand-written code at all, and called the discipline harness engineering: "the discipline shows up more in the scaffolding rather than the code." The prompt became a small part of a system, and the person's job stopped being operate the machine and became direct it.

Engineers kept climbing from here, into loops of agents and graphs of agents, and that ladder is real and worth reading. But it's theirs. Ours stops here, because the model can now think, see, and act, and what that exposes is the last thing on the list.

It cannot decide what you want.

The top of the ladder

Look at what every rung had in common. Each absorbed the one below it and kept it. Each was a workaround for something the model couldn't yet do, and each became redundant the moment it could. Every rule we ever wrote was about the partner.

Now look at the job market, because it only makes sense in that light. Those Indeed searches for "prompt engineer" that hit 144 per million in April 2023 settled at 20 to 30 and never rose again. Meanwhile the share of US job postings mentioning generative AI grew 3.5 times in the year to September 2024. And in January 2026, 7,359 US postings asked for prompt engineering as a skill, while out of 66,785 resumes in the same database, five people listed Prompt Engineer as their title. Five.

The protocol dissolved and the skill outlived it. That is only possible if the skill was never the protocol.

What it was, all along, is the thing the protocol was wrapped around. Knowing what a good result looks like before you ask for it. Saying the things you'd never say to a colleague, because they already know. Noticing what you assumed. Make it good was never a prompting error; it's an unfinished decision, and no rung on this ladder fixes it. The model can ask you clarifying questions now, and it will. It cannot decide for you what good means.

The gap between people who get real value from these models and people who don't has narrowed since 2023. It hasn't closed. It moved upstream of the prompt, out of technique and into the thirty seconds before you type.

The thirty seconds before the prompt

Five questions. Ask them first.

1. What does "done" look like? Could I recognise it? Could I recognise its opposite?
2. Who am I talking to? A brilliant stranger who knows almost everything and nothing about my situation. What would it need that I'd leave unsaid?
3. What am I assuming that I haven't said?
4. Have I asked what I'm missing before telling it what to do? A model handed a solution executes it. A model handed a problem may find a better one. Explore before you anchor.
5. What model am I actually using, and does it still need scaffolding?

That last one is the honest caveat. Everything above is true of the frontier. If your model is still 2023, your rung still applies: drop a rule, watch the output, keep whatever earns its keep. The principle was never abandon the techniques. It was know who you're talking to, and it holds on both ends of the ladder.

Beyond prompting

Every rule we wrote was about the model, and the model kept retiring them. The thing that survived is about you, and it doesn't retire when the next model ships.

We believe that gap is real; we see it every day in our own work. We haven't measured it, and we'd like to. Try the five questions on your next prompt and tell us what changed.