What Is Context Engineering? The Skill That Replaced Prompt Engineering

What Is Context Engineering, Exactly?

Context engineering is the practice of deciding what an AI tool knows before you ask it anything — the standing instructions, the background documents, the examples, and the memory it carries between chats. Prompt engineering is about the sentence you type. Context engineering is about everything already sitting in the room when you type it.

The term did not come from marketing. Anthropic’s engineering team published a detailed piece on effective context engineering in September 2025 describing it as the natural progression of prompt engineering, and introducing a phrase worth stealing: an AI model has an attention budget. Every token you add spends some of it, and spending it on filler leaves less for the thing you actually asked. That single idea is what context engineering is organised around. In a browser, it gives you four dials rather than one:

• Standing instructions — the rules that apply to every task in this workspace.
• Reference files — the documents the tool may consult.
• Examples — a couple of past outputs that show what “good” looks like here.
• Memory — what it carries between chats whether you asked it to or not.

Prompt engineering only ever touched the fifth thing: the sentence. Context engineering treats all five as one setup you design deliberately.

Notice what did not happen there. The second version is not a cleverer prompt. It is the same information, moved somewhere it does not have to be retyped. That is the whole idea, and it is why the term “prompt engineering” was always slightly misleading for office work.

Is prompt engineering dead?

No, and anyone telling you your prompt skills are worthless is overselling. Clear instructions still matter — they are just one dial among several now, and no longer the one with the most upside. Everything you know about writing good context into a request still applies. You are simply going to write it once instead of forty times.

The honest framing is that prompt engineering became a component of context engineering rather than a casualty of it. Writing a good standing-rules block is prompt writing. It just happens once, at setup, instead of every morning at 8:40.

The Stocked Desk Method

Most explanations of context engineering come from software teams and arrive full of retrieval pipelines and token budgets. Here is the office version, which fits on a napkin.

Picture a desk with two shelves above it. The shelves hold things you stock once and keep. The desk holds only what today’s job needs, and you clear it when the job is done. That asymmetry is the part most guides miss — two of your three layers are permanent, one is disposable, and mixing them up is where setups go wrong.

Shelf one: your standing rules

These are the instructions that are true on every task, forever: who the AI is acting as, who reads the output, the format you want, and the things it must never do. Most people write this once and are astonished how much drift disappears.

You are a [your job title] writing for [who reads this].

Always:
- Use plain business English. Short sentences.
- Keep to our house structure: [describe it, e.g. headline, one-line summary, three bullets, one action].
- Use British/American spelling: [pick one].

Never:
- Use the words [list your banned words here].
- Invent figures, dates, names or quotes. If a detail is not in the files I gave you, write "not specified" instead.
- Add a closing summary paragraph unless I ask for one.

When my request is ambiguous, ask me one clarifying question before writing.

Shelf two: your reference shelf

This is the small library the tool can consult: the style guide, the process document, the pricing sheet, two or three past pieces of work you were happy with. Those examples do a lot of quiet work — they are few-shot prompting made permanent.

Keep it lean. A reference shelf is a curated set, not an archive.

• Do stock: the current version of each document, cleaned of navigation junk and cover pages.
• Do stock: two or three genuinely good examples of the output you want.
• Do not stock: last year’s superseded policy, “just in case” background, or anything you would not hand a new starter.

The desk: today’s task only

Everything else goes in the chat itself — the file you are working on right now, the specific ask, the deadline. When the task is done, start a fresh chat inside the same project. The shelves stay stocked; the desk clears.

Where Your Three Controls Actually Live

Every major tool gives you the same three layers under different names. That is genuinely all you need to know to get started — once you can spot the pattern, switching tools costs you an afternoon rather than a relearning project.

Two things worth knowing before you pick. Claude’s Projects are available on the free plan, capped at five, so you can build a real setup without paying anything. Gemini’s Gems and Canvas are also on its free tier. ChatGPT is the one that gates the good version: you can use Custom Instructions free, but creating a Custom GPT with knowledge files needs a paid plan.

Which tool you pick matters far less than people expect. Context engineering is a set of habits about what you store and what you clear, and those habits survive a change of employer or a change of vendor. The buttons move; the three layers do not.

If you want the full walkthrough for your tool of choice, we have one for each: setting up Claude Projects, building a Custom GPT step by step, and creating Gemini Gems for recurring tasks.

A note on memory

Memory is the layer people forget they have. It carries facts about you between chats without you asking, which is useful right up until it quietly repeats something stale. Check what yours is holding — ChatGPT’s memory can be viewed and edited, and the same is true across the other tools.

Build Your First Shelf in 20 Minutes

Pick one task you do at least weekly. Not your hardest task — your most repetitive one. Weekly reports, client update emails, meeting summaries, job descriptions. Context engineering pays back on repetition, not difficulty, so a fiddly task you do once a year is the wrong place to start.

Create a Project, Custom GPT or Gem named after the task, not the tool. “Weekly depot update” beats “My AI assistant”. You will end up with several of these, and naming them by job makes the right one obvious six weeks from now.

Use the standing-rules prompt above. Aim for under 250 words. The most valuable lines are usually the negative ones — the words your organisation does not use, the format you never want, the invention rule. Vague guidance produces vague output.

Add three to six files, no more on the first pass. Before uploading anything, spend two minutes cleaning each one: delete cover pages, page headers, table-of-contents lists and anything superseded. Most of context engineering, for most professionals, turns out to be document hygiene rather than clever wording.

If a document is long and you are not sure what to strip, let the tool triage it first in a throwaway chat.

I am about to add the attached document to an AI project as reference material.

Tell me:
1. Which sections are genuine reference content worth keeping.
2. Which sections are navigation, boilerplate, cover pages or repetition I should delete.
3. Anything that looks out of date or contradicts another part of the same document.

Do not rewrite the document. Just list the sections by heading.

Give it last week’s real input and compare its output to what you actually sent. This is the step people skip, and it is the only one that tells you whether your shelves are stocked correctly or just stocked.

Using only the instructions and files in this project, produce [the deliverable] from the attached input.

Then, separately, list:
1. Anything you were unsure about because the files did not cover it.
2. Any instruction of mine that conflicted with another.
3. Any file you did not need at all.

That second half is the useful bit. The tool will tell you exactly which shelf item is missing, contradictory or dead weight — which is far faster than guessing.

Why Bigger Windows Don’t Fix This

A fair objection: if models now handle enormous amounts of text, why curate anything? Just upload everything and let it sort itself out.

Because capacity and attention are not the same thing. Chroma’s research team tested 18 models in a technical report on context rot published in July 2025 and found that performance degrades as input length grows — even on tasks deliberately kept simple, and even in the newest models at the time. The window tells you what fits. It does not promise the model will use all of it well.

In practice this shows up as a very specific frustration: the tool answers beautifully in a short chat, then in a long one starts ignoring a rule you set at the top. Nothing broke. You simply spent the attention budget on filler.

Which gives context engineering an unusual quality for a workplace skill — you improve results by removing things. Most professionals’ first instinct on a disappointing output is to add another paragraph of instruction or another supporting file. More often the fix is to delete the two files that were not pulling their weight.

When Your Context Goes Wrong

Four failures account for nearly everything that goes sideways in context engineering, and each has a specific fix rather than a general “try rephrasing”. Learning to tell them apart from the symptom is most of the skill.

The grounding rule is the highest-value line you can add anywhere in your setup. It converts a confident guess into an honest gap, which is the difference between a draft you can check and a draft you cannot trust. This is the same failure mode behind why AI tools invent details that sound right.

Answer using only the files and instructions in this project.

If something I ask about is not covered in those files, do not fill the gap from general knowledge. Instead write:

NOT IN PROVIDED FILES: [describe exactly what is missing]

Do this even if you are confident you know the answer.

One habit worth building: review your shelves once a quarter. Put it in the calendar. Stale context is the failure that creeps up slowly, because nothing looks broken — the answers simply drift further from what is currently true.

What This Cannot Fix

A well-stocked setup makes an AI tool consistent. It does not make it correct, and confusing those two is the mistake that gets people into real trouble at work. Consistency is easy to mistake for authority, particularly when the output arrives in your house format with your house vocabulary.

Consistency also has a downside worth naming. Once your rules are stored, a mistake in them repeats silently on every task rather than showing up once. Build the habit of checking output before your name goes on it, especially in the first fortnight.

Then there is the question of what you are uploading. A reference shelf is a pile of company documents sitting inside a third-party tool, and that deserves a moment’s thought rather than a shrug.

Check your organisation’s AI policy before your first upload, not after. On a personal free plan you are typically the only safeguard; on a company plan, an administrator has usually made some of these decisions for you already.

Which Setup Fits Your Work?

You do not need a project for everything. Most people overbuild at first, end up with eleven workspaces they never open, and conclude the whole approach was overhyped. Two questions sort it.

Start with one. Pick the task that annoys you most on a Monday morning, stock its two shelves properly, and run it for a fortnight before building a second. The people who get the most out of context engineering are not the ones with the most workspaces — they are the ones whose three or four workspaces are genuinely current.

And if it helps to know: the version of this skill that matters at work has almost nothing in common with the version being discussed by engineers building autonomous agents. Yours is a filing problem. Get the filing right and the prompting mostly takes care of itself.