Prompt engineer

What does a prompt engineer do?

A prompt engineer designs the instructions, context and checks that help a large language model (LLM) give useful, reliable answers.

The work covers writing prompts, connecting models to company data and testing outputs before an AI feature goes live.

In Singapore, the role sits in AI, data or product teams at banks, government agencies, logistics firms and software companies. Many employers use titles such as AI engineer. The closest official job code is artificial intelligence/machine learning engineer.

You work with product managers, software engineers, data teams, risk and compliance staff, and the business users who rely on the tool.

What the work involves

• Write and revise prompts and system instructions for a chatbot, summariser or classifier.
• Build test sets and score model answers to compare prompt versions.
• Connect models to company documents using retrieval-augmented generation (RAG).
• Call LLM APIs from Python and handle errors, cost and response time.
• Review wrong answers with users and trace each one to its cause.
• Record prompt versions, known limits and risks for product and compliance teams.

Pay in Singapore

Back to top

Skills a prompt engineer needs

Back to top

Prompt engineer job description

A Singapore organisation is hiring a prompt engineer to build and improve applications that use large language models.

You will write and test prompts, build retrieval pipelines and measure answer quality with evaluation sets. You will work with engineers, product owners and risk teams to move AI features from pilot to production.

Responsibilities

• Design, write and version prompts and system instructions for LLM-based features.
• Build retrieval pipelines that ground model answers in company documents.
• Create evaluation sets and scoring methods, and run them on every prompt or model change.
• Integrate LLM APIs into applications using Python, with sensible handling of errors, cost and delay.
• Test for harmful, biased, off-topic or leaked outputs, and for prompt injection.
• Analyse failed conversations and decide whether to fix the prompt, the data or the model.
• Check that personal data sent to models follows the Personal Data Protection Act (PDPA) and company policy.
• Document prompts, test results and limits so others can maintain and audit the system.
• Train business users on what the tool can and cannot do.

Requirements

• Diploma or degree in computer science, information technology, data science, linguistics or a related field, or an equivalent project portfolio.
• Typically 1 to 3 years in software development, data, natural language processing or applied AI work.
• Working Python and experience calling LLM APIs from at least one major provider.
• Experience with RAG and with frameworks such as LangChain or LlamaIndex.
• Ability to design tests for AI output and explain the results in clear writing.
• Awareness of PDPA duties and of the security risks in LLM applications.

Back to top

Prompt engineer career path

Prompt engineer has no single official career ladder, so steps differ by employer.

Common routes lead towards AI engineering, product or AI governance.

1. Junior prompt engineer or AI associate

   Entry

   You write and test prompts under guidance, build small evaluation sets and support the team's LLM features.

2. Prompt engineer

   Mid

   You own the prompts and evaluations for a product, build retrieval pipelines and work directly with users.

3. Senior prompt engineer or AI engineer

   Senior

   You design the application structure, set evaluation standards and review the work of others.

4. Lead AI engineer

   Lead

   You lead a team, set technical direction and join risk and security staff in release decisions.

The titles are ones commonly used in Singapore; employers may name the levels differently.

Related jobs and their pay

Back to top

Prompt engineer interview questions

We wrote these questions from the skills this job needs. They are not an official question bank.

1. How would you structure a prompt that extracts fields such as name, request type and urgency from customer emails?

   What a good answer covers

   • States the role, task and output format clearly, such as JSON with fixed keys
   • Includes a few examples, including an edge case
   • Says what to return when a field is missing
   • Tests the prompt on a labelled sample before release

2. What is RAG, and when would you choose it over fine-tuning a model?

   What a good answer covers

   • Explains retrieval of relevant documents added to the prompt
   • Prefers RAG when facts change often or sources must be cited
   • Notes fine-tuning suits style, format or narrow tasks
   • Mentions cost, maintenance and access control differences

3. How do you decide how to split documents into chunks for retrieval?

   What a good answer covers

   • Links chunk size to document type, such as policies, tables or FAQs
   • Keeps meaning together using headings or paragraph breaks
   • Describes overlap and its trade-off
   • Tests retrieval hit rate with real questions before settling

4. How would you build an evaluation set for a customer support assistant?

   What a good answer covers

   • Collects real or realistic questions across common and rare cases
   • Writes reference answers or scoring rules with domain staff
   • Covers refusals, unsafe requests and out-of-scope questions
   • Keeps a held-out set and updates it as usage changes

5. Model output varies between runs. How do you compare two prompt versions fairly?

   What a good answer covers

   • Runs each version on the same test set several times
   • Sets temperature and other settings explicitly
   • Uses clear scoring, with human review for subjective cases
   • Reports differences with the size of the sample, not a single example

6. Describe how you would call an LLM API from Python so it works reliably in production.

   What a good answer covers

   • Sets timeouts and retries with backoff for rate limits
   • Requests structured output and validates it
   • Logs inputs, outputs and token use with personal data protected
   • Keeps keys out of code

7. When would you use LangChain or LlamaIndex, and when would you write plain Python instead?

   What a good answer covers

   • Uses a framework for common parts such as loaders, retrievers and connectors
   • Uses plain code when the flow is simple and easier to debug
   • Names the cost of extra layers and version changes
   • Gives an example from a project

8. An assistant answers questions on HR policy and sometimes invents rules. How do you reduce this?

   What a good answer covers

   • Grounds answers in retrieved policy text and asks for citations
   • Tells the model to say when the answer is not in the sources
   • Improves retrieval before changing the prompt
   • Adds test cases for the invented rules and re-runs them

9. How do you defend an LLM application against prompt injection from user input or retrieved documents?

   What a good answer covers

   • Separates instructions from untrusted content
   • Limits what tools and data the model can reach
   • Checks outputs before actions run
   • Adds attack cases to the evaluation set

1. A chatbot scored well in testing, but users now report wrong answers in production. What do you do first?

   What a good answer covers

   • Collects and categorises the reported conversations
   • Compares real questions with the test set to find gaps
   • Checks whether documents, model version or settings changed
   • Adds the failures to the evaluation set before fixing

2. Your manager wants to move to a cheaper model next week. How do you check the effect?

   What a good answer covers

   • Runs the full evaluation set on both models
   • Reviews quality by task type, not only the average
   • Measures delay and cost per request
   • Recommends a staged rollout with a fallback

3. A business user asks for a prompt that never gets anything wrong. How do you respond?

   What a good answer covers

   • Explains that LLM output is probabilistic and errors can be reduced but not removed
   • Asks which errors matter most and sets measurable targets
   • Proposes human review for high-risk cases
   • Agrees how mistakes will be reported and fixed

1. Tell me about a prompt or AI feature of yours that failed. What did you change?

   What a good answer covers

   • Describes the failure with specifics
   • Shows how the cause was found
   • Explains the fix and how it was tested
   • States what you do differently now

2. Describe a time you explained the limits of an AI tool to colleagues without a technical background.

   What a good answer covers

   • Uses simple words and real examples
   • Shows respect for what the audience needs to do
   • Sets clear expectations on accuracy and review
   • Describes how the colleagues responded

1. Customer messages may contain NRIC numbers and phone numbers. How would you handle them when sending text to an external LLM API under the PDPA?

   What a good answer covers

   • Collects and sends only the data the task needs
   • Masks or removes identifiers before the call where possible
   • Checks the provider's terms on retention, training use and data location
   • Involves the company's data protection officer
   • Logs access and sets retention limits

2. Your assistant serves Singapore users who write in English, Singlish, Chinese, Malay and Tamil. How do you test and improve it?

   What a good answer covers

   • Builds test cases from real local phrasing and mixed-language messages
   • Measures quality per language, not only overall
   • Sets clear behaviour for unsupported languages
   • Reviews results with native speakers

Common questions about a prompt engineer job

Sources