You open a job board, type "prompt engineer", and get three listings. Type "AI engineer" and you get hundreds, some of which describe work you could do and some that assume you've trained models. So which one are you aiming for, and does the title even matter?
Here's the short version. Prompt engineering is the craft of getting reliable behaviour out of a model through instructions, examples and structure. AI engineering is the job of building and running software around that model, and it includes prompting as one part. Context engineering is the overlap: choosing what information reaches the model at each step. Most real jobs sit somewhere across the three, and the title is a weak signal. What you do all day is a strong one.
What does each role actually do all day?
These are archetypes, not job descriptions you'll find verbatim. Real roles blend them.
The prompt-focused role. Mostly work on behaviour. Writing and revising system prompts, building test sets of tricky inputs, comparing outputs across model versions, writing the style guide for what the assistant should and shouldn't say, reading failure transcripts and deciding whether the fix is a prompt change, an example, or a different approach altogether. Typical output: a prompt in version control, a test set, and a note on what changed and why. You might not write much code, but you'll read plenty and often write scripts to run evaluations. If you want a taste of this work, the prompt engineering learning path covers the skills in order.
The context-focused slice. The question is what the model sees. Which documents get retrieved, how they're chunked and ranked, what history is kept, what a tool returns and how much of it, what's stored in memory between sessions. Anthropic's engineering team defines the discipline as curating the smallest set of high-signal tokens that makes a good outcome likely (Effective context engineering for AI agents). Typical output: a retrieval pipeline, a budget for each part of the window, a compaction strategy. There's a worked method in context window budget.
The AI engineer. Owns the product feature end to end. Calls model APIs, wires up tools and retrieval, handles retries and timeouts, streams output, tracks cost and latency, builds the evaluation harness, and ships. Swyx's 2023 essay that popularised the title puts the line this way: AI engineers build on foundation models mostly via APIs, own product-specific data and evals, and don't need a research background (The Rise of the AI Engineer). The same essay argues "prompt engineer" is too narrow a name for software that mixes human-written code with LLMs. That was a 2023 opinion, and the ad-reading exercise below is how you check whether your market agrees.
A rough way to see the split: ask who gets paged when the feature is down at 2 a.m. The AI engineer does. The prompt-focused person gets pulled in the next morning when the answers got strange.
How are the titles drifting?
Honestly, nobody has clean numbers, so I'll be careful here.
What I can source: LinkedIn's Jobs on the Rise 2026 list puts "AI Engineer" first among 25 roles, with LangChain, retrieval-augmented generation and PyTorch listed as the most common skills for it (LinkedIn News). The list doesn't give a rank number or a growth figure for that role in the part I could read, so I won't quote one.
What I can't source: I found blog posts claiming the standalone "prompt engineer" title has dropped by 30 to 40 percent, and others claiming it grew several times over in 2025. They use different methods and I couldn't trace either to a primary dataset, so I'm leaving the numbers out. The consistent qualitative story across sources is that prompting skill is showing up as a requirement inside other titles. For the Indian market specifically, our prompt engineering salary in India post goes through the actual listings and pay ranges.
Do the check yourself, it takes fifteen minutes. Search your target job board for "prompt engineer", "LLM engineer" and "AI engineer". Open ten ads for each. For every ad, write down the three verbs that dominate the responsibilities (write, evaluate, build, deploy, monitor, research). That tally tells you more about your market than any headline.
Which job ad is this really? A decoder
Titles lie. Responsibility lines don't. Use this when reading an ad.
| If the ad says | The job is probably | Skill it will test |
|---|---|---|
| "Design and iterate prompts", "evaluate model outputs", "write guidelines" | Prompt-focused | Test sets, rubrics, writing, failure analysis |
| "Build RAG pipelines", "vector database", "retrieval quality" | AI engineering with a context slant | Chunking, ranking, evaluating retrieval |
| "Ship LLM features", "APIs", "latency and cost", "on-call" | AI engineering | Production code, observability, cost control |
| "Fine-tune", "train", "PyTorch", "papers" | ML engineering or research | Statistics, training loops, experiments |
| "Prompt engineer" with a Python and cloud stack list | AI engineering under an older title | Treat it as an AI engineer interview |
| "Prompt engineer" with no technical stack | Content or operations role | Writing and domain expertise |
The fourth row is worth noting: people often apply for ML roles thinking that's what "AI" means, when most application-layer roles don't touch training at all.
What skills does each path need?
Be honest about where you start, then close the gap that blocks the job you want.
For the prompt-focused path: precise writing, a habit of testing instead of trusting, and enough scripting to automate your test runs. The distinguishing skill is evaluation: building a set of cases, defining what "good" means, and noticing regressions. Start with the evaluation frameworks lesson.
For the AI engineer path: everything above, plus the unglamorous software parts. Retries, timeouts, structured output validation, logging, cost tracking, and security basics such as prompt injection. If you're a backend developer or test engineer, you already have the hard half. The India-focused career roadmap lays out a six-month plan for that route.
For the context slant: retrieval fundamentals, token budgeting, memory design. How RAG works is the base layer.
A trap in all three: collecting tool names. Listing ten chat products on a CV says you have used them. One small system that you built, tested against 20 cases and can show the failure rate of says you can do the job.
How do you pick a path?
Answer these in order and stop at the first yes.
- Can you write production code today? If yes, aim at AI engineer and treat prompting as a skill you already need. Pick a project that has retrieval, a tool call and an evaluation script.
- Is your edge domain knowledge (law, healthcare, finance, teaching) rather than code? Then the prompt-focused path is real for you: you can judge whether the output is right, which engineers often can't. Pair it with enough scripting to run your own evals. See prompt engineering for lawyers for what domain-led work looks like.
- Do you want to learn code but don't have it yet? Start prompt-focused for the quick wins, but plan the move to the engineering side within a year. Check the ad tally from the previous section to see whether your market agrees that the engineering side has more openings.
- Do you want to train models? That's a different track (ML engineering). Don't take an application-layer role expecting it.
Whichever you choose, build one portfolio piece that shows the loop: a problem, a baseline prompt, a measured failure, a change, a measured improvement. That single page is worth more than a certificate list, and it works for all three archetypes.
What to do this week
Run the fifteen-minute job-board check above. Then pick one ad you'd actually want and underline every skill in it you can't yet demonstrate. That underlined list is your plan. If you're starting from the beginning, what is prompt engineering and the learn section are the right first stops.
