Prompt Engineering

Context Engineering: How to Structure an LLM's Context

Context engineering means deciding what your AI reads before it answers. A 5-layer method, the mistakes that ruin results, and where to practice.

โดย Jérémie Chiari

15 กันยายน 2569 อ่าน 8 นาที

You write a clean, clear, well-turned prompt, and the answer comes back completely off target. Nine times out of ten, the problem isn't your sentence: it's everything the model had (or didn't have) in front of it at the moment it answered. Context engineering is the discipline of choosing that information, ordering it, and shaping it so the model works with the right raw material.

Here you'll see what a language model's context actually contains, how to structure it into five reusable layers, which mistakes quietly pollute it, and where to practice for real. No technical prerequisites: if you can hold a conversation with ChatGPT, Claude, or Gemini, you're ready. By the end, you'll have a context template to copy for your own tasks, and you'll finally understand why an overly long conversation almost always goes off the rails.

What exactly is context engineering?

Context engineering is deciding which information enters a model's context window, in what order, and in what form — before it even starts answering.

A language model remembers nothing between two requests. Every time you hit send, it rereads everything it's given: the starting instructions, the conversation history, attached files, tool results. That whole bundle is called the context window. It's measured in tokens, word fragments of roughly three to four characters (the word „context" weighs one or two tokens, a page of text around 400).

That window is large, but it's finite. Recent Claude models accept 200,000 tokens, or several hundred pages, according to the official documentation on context windows. Large doesn't mean unlimited, and more to the point: the more you fill the window with useless material, the more the decisive information gets diluted.

	Prompt engineering	Context engineering
Focus	how you phrase your request	everything the model reads
Scope	one message	a session, a project, an agent
Typical question	„how do I word this?"	„what does the model actually need?"
Example	adding „answer in 5 bullet points"	supplying the style guide and 3 already-published articles

The two complement each other. Prompt engineering optimizes the request; context engineering prepares the material. If you're just starting out, the complete method for writing a good prompt is still the foundation everything else rests on.

What an LLM's context really contains

Context isn't just your question: it stacks at least six blocks of information, and your question is often the smallest of them.

System instructions. The invisible frame set by the tool or by you: role, tone, prohibitions, response language. In ChatGPT these are the custom instructions; in Claude, the project instructions.
Conversation history. Every previous message, yours and the model's, is sent back on every turn. A 40-message exchange gets paid for again with each new question.
Documents and data you provide. PDFs, spreadsheets, a pasted web page, a database extract. This is where the volume explodes fastest.
Examples. Two or three samples of what you're after beat ten lines of adjectives. That's the whole principle behind few-shot prompting.
Tools and their results. Web search, code execution, API calls: every result comes back into the window, sometimes running to thousands of tokens.
Your actual request. Often a single sentence, buried at the end of all the rest.

Out of 20,000 tokens of context, your final question might weigh 40. That's exactly why a model that „isn't listening" is usually a badly fed model, not a broken one. The Anthropic team describes context as a finite resource you have to budget, just like time or money.

Structuring your context: the five-layer method

Always arrange your context in the same order: role and goal, constraints, reference data, examples, then task and output format.

That order isn't decorative. It puts the frame before the material, and the action instruction right before generation — the place the model reads last.

Layer 1: role and goal. Who's speaking, to whom, for what purpose. Two sentences is enough. „You're the content lead at an accounting firm serving small trade businesses. Goal: make tax topics understandable without oversimplifying them."

Layer 2: constraints. Length, tone, banned vocabulary, legal obligations, target audience. Write them as a list, not a paragraph: one constraint per line gets respected far more often than one hidden inside a sentence.

Layer 3: reference data. The facts the model can't guess: your prices, your figures, your client list, the source text to rework. Fence them off clearly, for example between <data> and </data> tags, and tell the model explicitly that it must not invent anything outside them.

Layer 4: examples. One or two representative samples of what you call „good." If you have a bad one, show that too and explain what's wrong with it.

Layer 5: task and output format. The exact action, then the expected shape: word count, structure, table, JSON, headings. Always finish here.

Here's what the result looks like, in short form:

You write a newsletter for independent florists.

Constraints: 250 words max, casual and direct, no marketing jargon, one idea per paragraph.

Data: our January offer (bouquet subscription at €24 a month, delivery included in Lyon, cancel anytime).

Example of the tone we want: „This month we got ranunculus in. They last eight days if you change the water every couple of days."

Task: write the February edition. Format: one headline, three paragraphs, a closing sentence with a call to action.

That block becomes a template. From then on you only rewrite layers 3 and 5, which cuts the time spent preparing each request by two thirds.

The five mistakes that pollute a context

Most bad answers come from a context that's overloaded, contradictory, or badly ordered — not from a model that isn't smart enough.

Massive copy-paste. Pasting an 80-page report to get two chapters summarized is like asking the model to find a key in an attic. Pull out the useful passages, or split it into several requests.
Stale context. You said „formal tone" at message 3, then „loosen up" at message 25. Both instructions coexist in the window and the model arbitrates on its own. When you change your mind, rewrite the full instruction instead of patching it in small touches.
Key information buried in the middle. What matters most goes at the start (the frame) or at the end (the task). The belly of the context is where attention slackens.
No output format requested. Without a format instruction, the model picks its own, and it changes with every attempt. It's one of the recurring faults covered in the prompting mistakes that ruin your results.
The endless conversation. After 50 messages, your context is mostly back-and-forth, abandoned attempts, and apologies from the model. Open a new conversation and start again from your updated five-layer block. That's often the move that unblocks a session going in circles.

Where to practice context engineering: three routes

Three options depending on your budget and level: a guided program with the tools included, the project features of the AI you already use, or the vendors' technical consoles.

Route 1: a guided program, tools included (Skilzy)

This is the fastest path if you're starting from zero. The Skilzy context engineering program has you build your own context blocks on real cases (product page, newsletter, customer support, automation), with feedback and reusable templates. Skilzy is a French platform with more than 15 programs, from image and video creation to automation with n8n, and its built-in AI Lab gives you access to the real tools with credits included: you practice without stacking up three €20-a-month subscriptions.

Access starts at €29.90 a month, no commitment. Two state-recognized certifications eligible for training funding are available (RS7439 in AI content marketing, RS6792 in AI and sales). To try before you pay, the discovery demo runs 7 days with 1 image, 1 video, 1 music track, and 10 messages, no credit card.

Route 2: your AI's projects and custom instructions

Free, or included in your current subscription. Projects in ChatGPT and Claude, like Gemini's Gems, exist for exactly this: you drop in your system instructions, reference documents, and examples once, and every new conversation starts with that context already loaded. Ideal for testing the five-layer method today. Limitation: you're working without outside feedback, and it takes several weeks of trial and error to figure out on your own what works.

Route 3: official documentation and vendor consoles

The most precise path, and the most demanding. The documentation from Anthropic, OpenAI, and Google covers context window management, caching for long preambles, and message structuring. The developer consoles let you count tokens and compare two versions of the same context. Useful the moment you want to scale up, indigestible if you've never touched an API.

Key takeaways

A good result rarely comes from a magic sentence. It comes from a clean context: a clear role, listed constraints, verified data, two examples, a task, and a format. Build that block once, reuse it, and clean it out as soon as the conversation gets heavy. You'll spend less time fixing vague answers than preparing what you give the model to read — and that's exactly the right balance.

คำถามที่พบบ่อย

What's the difference between prompt engineering and context engineering?

Prompt engineering is about how you phrase a specific request: action verb, level of detail, expected format. Context engineering is about everything the model reads before answering: system instructions, history, documents, examples, tool results. The first works on one message, the second on a whole session or project. They're complementary: a perfect prompt sitting on a polluted context still produces a bad answer.

What's a token and why does it matter?

A token is a word fragment, usually three to four characters. Models count their context in tokens, not words: a page of text weighs roughly 400 tokens. It matters because the context window is limited (200,000 tokens for recent Claude models) and because API billing is per token. The more useless context you send, the more you pay and the more the important information gets diluted.

Should you put everything in the context to get a better answer?

No, that's the most common mistake. An overloaded context degrades quality: the model has to sort through it itself, and sometimes it sorts badly. The good practice is to provide the minimum that's sufficient: the facts the model can't guess, one or two representative examples, and the format instruction. If you're unsure whether to include a document, ask yourself whether you'd use it yourself to do the task.

Why does a long conversation with an AI end up going off the rails?

Because the entire history is reread with every message. After a few dozen exchanges, the window contains contradicted instructions, abandoned directions, and rewordings that cancel each other out. The model then arbitrates between incompatible instructions. The fix is simple: open a new conversation and paste in a clean, up-to-date context block with the final version of your instructions.

Can you practice context engineering for free?

Yes, partly. The Projects features in ChatGPT and Claude, Gemini's Gems, and the vendors' official documentation are all available at no extra cost if you already have an account. On the Skilzy side, the discovery demo gives you 7 days of access with 1 image, 1 video, 1 music track, and 10 messages, no credit card, so you can see what the programs look like before subscribing at €29.90 a month.