The AI Skill Worth Learning Now: Context Engineering

There's a comforting story going around that "prompt engineering" was the hot new career. As an EM who hires for these skills and uses the tools every day, I'll be blunt: it was never really a job — and the models got good enough to make the title disappear. What's actually sticking, and what's worth your time, is context engineering.

Prompt engineering is what you say to the model. Context engineering is everything the model can see when it answers.

That sounds like a small distinction. It changes how you work.

Why the title died but the skill didn't

The "$200K prompt engineer" headlines peaked back in 2023. Two years later that standalone role has mostly evaporated — not because the skill stopped mattering, but because it spread into every job and the models started asking their own clarifying questions. AI literacy is now one of the fastest-growing skills employers ask for, yet almost nobody hires a "prompt engineer" anymore. That's the pattern to internalize for your own career: AI rarely deletes the work, it deletes the job description wrapped around it.

Jensen Huang put the career math plainly: "It is most likely that most people will lose their job to somebody who uses AI," not to AI itself. His distinction is the one that matters — your tasks, and the job those tasks add up to, are related but not the same. Context engineering is how you become the person on the winning side of that sentence.

What context engineering actually is

The moment you build anything past a one-off chat — an agent, a workflow, a Claude Project — the prompt is a tiny slice of what decides the output. The real levers are the system instructions, the tools you expose, the examples you show, what you retrieve, and how much history you carry between turns. Anthropic defines it as curating and maintaining the optimal set of tokens the model sees during inference, and the goal it sets is worth memorizing: find the smallest set of high-signal tokens that gets the outcome you want. More context isn't better; the right context is.

Andrej Karpathy — who popularized the term — has the mental model I keep coming back to: picture the model as a computer and the window as its RAM. You're the one loading that memory before each run, so put in only what the task needs and leave the rest out. There's even a named failure mode: context rot. Research Anthropic cites shows that as the token count climbs, the model's ability to accurately recall what's in the window drops. Cram in everything and it gets measurably worse at finding what matters.

The six levers you actually control

"Everything the model sees" is true but too vague to act on. Break it into the parts you control and you can run the same checklist before any task that matters. There are six.

• Instructions. The role the model is playing and the guardrails it works inside — name the role, then spell out what it must always do and what it must never do.
• Knowledge. The real material: your actual docs, numbers, and data — the artifacts themselves, pasted in, not a summary of what they say.
• Examples. A sample or two of the output you'd call great. Anthropic's guidance is to curate a few diverse, canonical examples rather than an exhaustive edge-case list — one good example carries more than a paragraph of instruction.
• History & memory. A deliberate call on which prior turns to keep and which to let go. Stale, off-topic history quietly drags every later answer down; for longer work, park durable facts in a memory file instead of the live window.
• Tools. The actions and surfaces the model can touch — files, the web, your calendar, an API. Keep the set small and each tool unambiguous; Anthropic warns that bloated, overlapping tool sets create decision points an agent handles no better than a human would.
• Format. The shape you want back — a table, a short doc, a diff, five bullets. State it before the ask, and reach for delimiters or a schema when the structure has to be exact.

Track down a disappointing answer and it's almost always one of these six that was missing. Covering all six is what separates getting a reply from getting a result.

The scaffold I paste at the top of anything that matters

Keep this in a note and fill in the blanks. What jumps out is how little of it is an actual "question" — that's deliberate. Lay the table first and the ask almost writes itself.

ROLE: You are my [role/expert, e.g. senior backend engineer].
GOAL: I need [the exact outcome, e.g. a migration plan for this service].
CONTEXT (the real material):
  - [paste the actual code, doc, data, or error here]
  - About how I work: [paste your "about-me" file]
  - Who this is for: [audience / teammate / user]
RULES:
  - Do: [what a great answer must include]
  - Don't: [what to avoid — no hype, no invented APIs]
  - If you're missing something you need, ask before you guess.
EXAMPLE OF GREAT (optional, powerful):
  - [paste one example of the output you want]
FORMAT: Give it to me as [a table / 5 bullets / a diff].

The four moves on one real prompt

Watch the levers do their job on an actual support-reply agent. Before — everything, just in case:

[full 12-page product manual pasted]
[last 40 messages of chat history]
[every reply template we own]
[3 tool descriptions, two of them near-identical]
Answer the customer.

After — the same job, curated:

SYSTEM: You answer billing questions. Warm, two sentences max.
RETRIEVED: [only the refund-policy section, pulled by the customer's question]
EXAMPLE: Q: "Charged twice?" → A: "I see the duplicate — refunding the extra
now, it lands in 3-5 days. Sorry for the scare."
TOOLS: lookup_charge (find a charge by email) · issue_refund (reverse one charge)
Customer says: [message]

Cut the manual and stale history, retrieved one section on demand, taught with one example, and named two tools you can't confuse. Instructions, knowledge, examples, tools, format — all six levers, set deliberately.

Five habits that separate the pros

• Tight beats big. Piling on context doesn't help — it dilutes. Feed the task what it needs and stop there; context rot is the tax on everything extra.
• Show, don't adjective. A single real example does more than ten adjectives like "professional" or "clean." When you can, paste the example instead of describing the vibe.
• Position counts. Models weight the opening and the close of the window most, so drop your highest-value context right next to the ask — not buried in the middle.
• Fix the context, not the wording. When the output is off, don't rephrase the question. Find the missing piece, add it, rerun. Building that reflex is the skill.
• Reset often. Once a thread turns long and cluttered, don't fight it — spin up a fresh chat and drop in a clean context block. A messy window dulls even a sharp model.

Three drills for this week

1. The rebuild. Grab a bad answer from earlier today. Instead of rephrasing, feed it the one fact it was missing and rerun — then notice precisely what flipped it.
2. The golden file. Draft your "about-me" doc — role, goals, stack, voice, rules — and lean on it for a week, tightening it each time you catch yourself repeating context you've already given.
3. The trim. Grab a bloated thread and halve its context. If the answer survives, you just found out how much the model was carrying for nothing.

Where to go deeper

• Anthropic — Effective Context Engineering for AI Agents. Canonical reference, written by the people building Claude — the sharpest breakdown of each lever and how to keep the window lean. Start here.
• Andrej Karpathy's take. From the engineer who put the term on the map; his working-memory analogy is what makes the whole idea click.
• The Prompting Guide — Context Engineering. A free, no-jargon walkthrough that goes component by component with worked examples. Good for practicing.
• Anthropic Cookbook — Context Engineering with tools. Hands-on code for when you graduate to building agents and have to manage tools, memory, and long context. The next step once the basics land.

Do one thing today

Open the Anthropic piece, then take drill #1 straight into something real on your plate at work. That loop — a little learning, applied the same day on the job — is what makes it stick. You won't get there by reading; you get there by repetition. Run it a handful of times and you'll own the skill the market is actually paying for.

Sources

• Anthropic — Effective context engineering for AI agents
• Fortune — the $200K prompt engineering role is now obsolete
• Fortune — Jensen Huang on AI replacing tasks, not jobs
• Andrej Karpathy — context as the model's working memory
• Prompt Engineering Guide — Context Engineering
• Anthropic Cookbook — context engineering with tools