Context Engineering: The Skill to Learn After Prompt Engineering

This is the guide from my CONTEXT reel. In the reel I said that prompt engineering alone is no longer enough, and that companies now need one more skill. Here is the full version: what context engineering is, a 6-step learning list, one beginner project, and free resources for each step.

Prompt vs context in one minute

Prompt = what you tell the AI. The instruction.
Context = what the AI can see when it answers: documents, data, earlier conversations and tools.

Prompt engineering still matters. Context engineering is the next layer on top of it. Anthropic’s engineering team describes it as choosing and maintaining the right set of information in the model’s limited working space, and notes that a model recalls things less accurately as that space fills up. So more is not always better.

A simple example

You ask a support bot, “Where is my order?” A perfect prompt does not help if the bot cannot see your order. Someone has to decide which systems the bot reads, how much of your history it gets, what it is not allowed to see, and how to keep that information current. Deciding what the AI sees, how much it sees and what it must not see is context engineering.

What a context engineer does

Decides what information the AI gets for each task. Too little and it guesses. Too much and it gets confused and slow.
Shapes company data so the AI retrieves the right evidence.
Works out how data moves between systems such as the CRM, documents and databases.
Sets permission rules, so the AI never shows one person another person’s data.
Helps turn a demo into a system that works reliably.

The honest part about the job title

Do not expect many job posts titled “Context Engineer”. In a TechRadar Pro opinion article dated 5 August 2026, Tobie Morgan Hitchcock, CEO of the database company SurrealDB, writes that some organisations may hire for that exact title but many will not. He says the work is more likely to appear inside roles such as AI engineer, agent engineer, AI platform engineer, applied AI engineer or data engineer. So learn the skill, and search for those titles.

The learning list: 6 steps, in order

Before you start: most of the free resources below expect basic Python. If you have never coded, learn Python basics first, or do the project in its no-code form (explained in the project section). The order and the “you are done when” lines are my own suggestions. Go at your own pace; nobody can honestly promise you a fixed number of days.

Step 1: Retrieval (RAG)

What it is: finding the few right pieces of your documents and giving only those to the AI, instead of everything.
Learn: splitting documents into chunks, embeddings (turning text into numbers so similar meanings sit close together), vector search, and re-ranking the results.
You are done when: you can ask a question over your own 20 to 30 documents and show which chunk the answer came from.
Free resource: Hugging Face cookbook: Advanced RAG, a worked notebook that goes through chunking, embeddings, vector search and re-ranking with open-source tools.

Step 2: Data pipelines

What it is: getting data from where it lives (sheets, a CRM, a database) into a form the AI can use, and keeping it fresh.
Learn: basic SQL, loading and cleaning data, and running an update on a schedule.
You are done when: a change in your source data (for example an order status) shows up in the AI’s answer without you copying anything by hand.
Free resource: Data Engineering Zoomcamp by DataTalks.Club. Its page describes it as a free 9-week course and says you can start anytime; the certificate needs a live cohort. It is bigger than a beginner needs, so start with the data ingestion and workflow parts.

Step 3: Agents and tools

What it is: letting the AI call a tool (look up an order, search a file) to fetch information when it needs it. MCP, the Model Context Protocol, is described on its official site as an open-source standard for connecting AI applications to external systems.
Learn: how to describe a tool clearly, what a tool should return, and why a small set of clear tools works better than a long list of overlapping ones (a point Anthropic’s article makes).
You are done when: your assistant can call one tool you wrote, such as “get order by ID”, and use the result in its answer.
Free resources: Hugging Face AI Agents Course (free, with a free certificate and no deadline; needs basic Python), Introduction to Model Context Protocol on Anthropic Academy (free registration; needs Python), and the official MCP introduction.

Step 4: Memory

What it is: deciding what to keep from earlier in a conversation, what to save for next time, and what to drop.
Learn: short-term memory (inside one chat) versus long-term memory (across chats), summarising a long chat before it gets too big, and having the AI keep notes outside the chat that it reads back later.
You are done when: your assistant remembers the order ID you gave three messages ago, and does not carry it into the next customer’s chat.
Free resources: LangChain docs: Memory overview for short-term versus long-term memory, and Anthropic’s article in the last section for summarising and note-taking.

Step 5: Access control

What it is: who is allowed to see what. The AI should only ever receive data that the person asking is allowed to see.
Learn: filtering data by user before it reaches the AI, and putting the rule in the database rather than in the prompt. A prompt that says “do not reveal other customers’ orders” is a request, not a lock.
You are done when: user A asks for user B’s order and the system has nothing to show, because that row never reached the AI.
Free resource: Supabase docs: RAG with Permissions, which shows a database rule (Row Level Security) that returns only the document chunks a user owns. The idea carries over to any database.

Step 6: Evaluation

What it is: testing whether a change to the context made answers better or worse, with numbers instead of a feeling.
Learn: writing a fixed set of test questions with expected answers, scoring every version against the same set, and changing one thing at a time.
You are done when: you have a small table that shows the score before and after each change you made.
Free resources: Hugging Face cookbook: RAG Evaluation and the Claude docs page Define success criteria and build evaluations.

The project: a “Where is my order?” bot

This is the example from the reel, turned into a project. It is small, but it touches all 6 steps. Use made-up data only. Never put real customer names, phone numbers or orders into a practice project.

What you need: a sheet with about 20 fake orders, a one-page delivery and returns policy that you write yourself, and any AI assistant. Give the sheet these columns:

Column	Example
order_id	ORD-1007
customer_phone	a fake 10-digit number
item	Blue water bottle
status	Packed, Shipped, Out for delivery, Delivered
expected_delivery	a date
last_update	date and time of the last status change

Write 20 test questions first. For example “Where is order ORD-1007?”, “Can I return it?”, and a few trick ones such as an order ID that does not exist, or a customer asking about someone else’s order. Write the correct answer next to each. This is your evaluation set.
Version 0: prompt only. Give the AI the instruction below and no data. Ask your 20 questions and score them. It will fail or guess on almost all order questions. That is the point of the reel in one test.
Version 1: everything. Paste the whole sheet and the policy into the chat. Score again. It works at 20 orders. Note the problem: every customer’s data is now in front of the AI, and this approach will not fit when there are 20,000 orders.
Version 2: only the right row. Look up the one order the customer asked about and give the AI only that row, plus only the policy paragraph that matches the question. With no code you do the lookup by hand. With code, this becomes a small “get order by ID” tool. Score again.
Version 3: add the permission rule. The lookup returns a row only when the phone number of the person asking matches the phone number on the order. Re-run your trick questions and check that nothing leaks.
Version 4: add memory and freshness. The bot should remember the order ID within one chat and forget it for the next customer. Then change a status in the sheet and check that the next answer shows the new status.
Write it up. One page: the score for each version, what you changed, and what you learned.

Use the same instruction for every version, so that the only thing you change is the context:

You are a support assistant for a small online store.
Answer only from the order data and policy text given to you in this chat.
If the order is not in the data you were given, say you cannot find it and ask for the order ID.
Never guess a delivery date. Keep answers under 3 sentences.

What to show in your portfolio

The score table for versions 0 to 4 on the same 20 questions.
One wrong answer from version 0 next to the right answer from version 2 for the same question.
The permission test: the question, and proof that the other customer’s row never reached the AI.
A simple diagram of what the AI sees at each step.
What you would do next with real data. If you coded it, add the link to the code.

If you run a small business, the same exercise is a useful check before you put any AI bot in front of customers: what will it see, how fresh is that data, and what must it never see?

All the free resources in one place

I checked each link on 11 November 2026. Courses change, so confirm on the page that it is still free before you sign up.

Start here: Anthropic: Effective context engineering for AI agents (29 September 2025). One article that explains the idea, why too much context hurts, and how to handle long tasks with summaries and notes.
Retrieval: Hugging Face cookbook: Advanced RAG. Also Building with the Claude API on Anthropic Academy (free registration), which includes a section on RAG with chunking and embeddings.
Data pipelines: Data Engineering Zoomcamp.
Agents and tools: Hugging Face AI Agents Course, Introduction to Model Context Protocol, and the official MCP introduction. The full Anthropic Academy catalogue is here.
Memory: LangChain docs: Memory overview.
Access control: Supabase docs: RAG with Permissions.
Evaluation: Hugging Face cookbook: RAG Evaluation and Claude docs: Define success criteria and build evaluations.

Source

TechRadar Pro, 5 August 2026: Why context engineering is AI’s next hiring challenge, an opinion article by Tobie Morgan Hitchcock. The learning list, the “you are done when” checks and the project are my own suggestions. This page is general career information, not a promise of a job.