What does a prompt engineer do?

What the work involves day to day

The visible part of the job is writing instructions for a model, especially system prompts, and revising them until the output is reliable. Anthropic defines prompt engineering as methods for writing and organizing LLM instructions to get the best outcomes. In practice you also decide which examples, rules, output formats and reference data go into the request, because each of them changes what the model does.

The bigger part is measurement. OpenAI's evaluation guide sets out the loop: define what success means, collect a dataset, choose metrics, then run and compare evals. It warns against judging a system on whether it seems to work, and it says automated scores should be checked against human feedback. A prompt change with no eval behind it is a guess.

The rest is comparison and documentation. When a new model version ships, you rerun the same eval set on it and on the current model before switching. You write down which prompt version shipped, why it won, and which cases still fail, so the next person does not repeat your tests.

Is prompt engineering still a standalone job?

Mostly no. Snowflake's 2026 guide says the standalone prompt engineer title has declined from its early peak and that prompt work has moved into AI engineering, data engineering, application development and product roles. Its reason is practical. Prompts interact with retrieved data, tool definitions, output schemas and evaluation logic, so the people who build those parts usually own the prompts as well.

Anthropic now describes context engineering as the natural next step after prompt engineering. Agents run over many turns, so the work covers everything the model sees: system instructions, tools, external data and message history. IEEE Spectrum reported in 2024 that automated prompt optimizers often beat hand-tuned prompts when given examples and a scoring metric. Building that scoring metric is still human work.

Where the skill shows up in job ads

Coursera's guide to prompt engineering jobs notes that a search for prompt engineer jobs may return only a handful of postings with that exact title, and that many roles include prompt work daily without naming it. On AutomationRoles.ai you will find it inside AI engineer, AI agent developer, conversational AI and AI automation roles. Look for duties such as building evals, comparing models, writing system prompts and maintaining guardrails.

Domain knowledge is the part models cannot supply. Snowflake points out that a capable model still does not know your company's definition of revenue, when a decision needs approval, or which source to trust when two systems disagree. People who can write those rules down precisely and then test them have a lasting advantage, whatever their job title says.

Entry-level reality

Do not count on prompt engineer as a first job. Anthropic's own report on its workforce says it has been hiring a smaller share of junior technical staff, and hiring managers put more weight on autonomy and judgment than on specific technical expertise. A safer plan is to learn prompt and eval work as part of a broader role, such as software engineering, data science or product, where it adds to skills employers already hire for.

Pay

USD a year, before tax. US figures from BLS OEWS May 2025, moved to 2026 with the Employment Cost Index. BLS has no prompt engineer or AI series, so this is the nearest match: the software developers series, which covers the whole occupation and not AI specialists alone. State and the depth of engineering in the role move pay the most. Refer to the LLM & Prompt Engineers salary guide for figures from live adverts.

What you need

• Clear technical writingYou need to write instructions, rules and edge cases precisely enough that a model and a colleague read them the same way.
• Evaluation designDefine success criteria, build a test dataset that looks like real traffic, choose metrics, and check automated grading against human judgment.
• ProgrammingCoding skills help you call model APIs, run eval sets in batches and compare results.
• Context assemblyKnow how retrieved documents, tool definitions, output schemas and conversation history affect a model's behavior, along with the prompt text itself.
• Domain knowledgeBusiness rules, approval steps and data definitions have to come from someone who understands the work being automated.
• Documentation habitsVersion your prompts, record why each change shipped, and keep a list of known failure cases.

How to get there

1. Year 0-2Build a base roleStart in software engineering, data analysis, data science or product, where prompt work adds to skills employers already hire for.
2. Year 1-2Learn evals on real tasksPick a task at work, build a scored test set, and show measured gains from prompt or model changes.
3. Year 2-4Own an LLM featureTake responsibility for prompts, retrieval, output format and regression tests on one production feature.
4. Year 3-5Move to an AI engineering titleApply for AI engineer, agent developer or conversational AI roles that list prompt and eval duties.
5. Year 5+Specialize or leadGo deeper into evaluation and model quality, or lead the team that builds and tests LLM features.

Open now on AutomationRoles.ai

Prompt EngineerParadise Architectural PANELS & STEEL · San Francisco, California

→

Full Stack LLM EngineerCerebras · Toronto, Ontario

→

LLM EngineerTraining Basket · Noida, Uttar Pradesh

→

LLM EngineerZS · Toronto, Ontario

→

Sources

1. Anthropic: Effective context engineering for AI agents
2. OpenAI: Evaluation best practices
3. OpenAI: Working with evals
4. Snowflake: Prompt engineering for production AI
5. IEEE Spectrum: AI prompt engineering is dead
6. Coursera: Prompt engineering jobs career guide
7. Anthropic: How employment is changing at Anthropic
8. BLS Occupational Outlook Handbook: Software developers