Generative AI & Prompt Engineering

Understand generative models and select a useful application problem.

• Language, image and multimodal model concepts.
• Tokens, context windows, generation and model limitations.
• Pretrained models, inference and adaptation concepts.
• Use-case selection, user needs, success criteria and cost considerations.

Practical exerciseWrite an assistant project brief with a defined user, available information and measurable success criteria.

Understand what language models can do and how to express a task clearly.

• Model capabilities, limitations, context and variability.
• Task, audience, context, constraints and expected output.
• Separating instructions from reference material.
• Recognising when a task needs a tool, a source or a human decision.

Practical exerciseRewrite an unclear workplace request into a prompt with a defined audience and an assessable output.

Choose an instruction pattern that fits the task and compare the results.

• Zero-shot and few-shot prompting with relevant examples.
• Summarisation, classification, extraction and rewriting tasks.
• Task decomposition, checklists and stepwise workflows.
• Managing ambiguity, asking useful clarification questions and stating assumptions.

Practical exerciseCreate prompts for three different tasks and explain which examples and constraints improved the output.

Make responses traceable to the information supplied to the model.

• Selecting relevant reference material and fitting it within the context.
• Quoting, summarising and preserving the meaning of a source.
• Source attribution and handling conflicting or missing information.
• Distinguishing supported answers from speculation.

Practical exerciseBuild a document question-answering prompt that cites supplied sections and acknowledges unsupported questions.

Design consistent outputs that people and simple workflows can use.

• Tables, schemas, JSON concepts and fixed response formats.
• Variables, placeholders and reusable prompt templates.
• Validation rules, missing fields and malformed-output handling.
• Versioning prompts and keeping examples aligned with the task.

Practical exerciseCreate a reusable extraction template and check its outputs against required fields and allowed values.

Apply prompting to practical tasks while checking the quality of the final result.

• Audience-aware writing, tone, editing and content repurposing.
• Research questions, source review and fact-checking workflows.
• Comparisons, categorisation and summarising supplied data.
• Reviewing calculations, quotations and unsupported claims.

Practical exerciseCreate a research brief and audience-specific summary from a supplied set of documents.

Improve prompts through repeatable checks instead of judging a single good response.

• Success criteria for correctness, completeness, relevance and style.
• Representative examples, difficult cases and regression checks.
• Comparing prompt versions and recording failure patterns.
• Human review, scoring rubrics and variation across repeated runs.

Practical exerciseBuild a small evaluation set, compare two prompt versions and document the improvement and remaining failures.

Use AI responsibly when a task involves sensitive information or actions outside the conversation.

• Data minimisation, confidential information and redaction.
• Prompt injection and treating external text as untrusted input.
• Tool-use boundaries, confirmation steps and human approval.
• Bias, harmful assumptions and escalation when confidence is insufficient.

Practical exerciseReview a workplace AI workflow and add checks for sensitive data, unsafe instructions and unsupported actions.

Package useful prompts with evidence that another person can reproduce and evaluate them.

• A clearly scoped workplace problem and user instructions.
• Reusable templates with sample inputs and expected output criteria.
• Evaluation results, limitations and a maintenance checklist.
• Explaining prompt decisions, failures and improvements in a project discussion.

Practical exercisePresent a prompt portfolio with tested examples, an evaluation rubric and a short usage guide.

Build the foundations for connecting a model to an application workflow.

• Python functions, data structures, HTTP and JSON.
• Messages, instructions, context and structured outputs.
• Server-side secret handling, input validation and response parsing.
• Timeouts, rate limits, bounded retries and useful error messages.

Practical exerciseCreate a small model-backed workflow that validates input and handles unsuccessful requests clearly.

Define how an assistant should respond and what it should do when information is missing.

• System instructions, task prompts and relevant examples.
• Structured response schemas and deterministic application validation.
• Conversation state, context selection and token budgets.
• Uncertainty, clarification and fallback behaviour.

Practical exerciseBuild a structured assistant response and test incomplete, ambiguous and out-of-scope requests.

Ground an assistant in a document collection rather than relying only on model memory.

• Document loading, cleaning, chunking and metadata.
• Embeddings, similarity search and vector-store concepts.
• Retrieval, ranking, prompt assembly and source attribution.
• Access-aware retrieval, missing evidence and stale documents.

Practical exerciseBuild a document-based question-answering flow that identifies supporting passages and handles missing answers.

Connect model decisions to controlled application steps.

• Tool/function schemas, arguments and validation.
• Multi-step workflows, state and termination conditions.
• Read-only versus side-effecting actions and human approval.
• Agent failure modes, loop limits and Model Context Protocol concepts.

Practical exerciseDesign an assistant that looks up supplied information and prepares an action for review before execution.

Explore how additional input types and adaptation choices change an application.

• Text, image and document inputs; extraction quality and limitations.
• Speech-to-text and text-to-speech workflow concepts.
• Prompting, retrieval and fine-tuning: choosing the appropriate approach.
• Dataset preparation, consent and evaluation before adaptation.

Practical exercisePrototype a document-understanding workflow and compare its extracted fields with a reviewed reference.

Measure quality and design for reliable operation before exposing an assistant to users.

• Task-specific evaluation sets, groundedness and answer quality.
• Prompt injection, data leakage, access checks and output handling.
• Latency, token usage, cost, logging and feedback.
• Versioning, monitoring, rollback and human escalation.

Practical exerciseEvaluate an assistant against representative and adversarial inputs, then document deployment and monitoring checks.

Deliver a focused assistant application and explain how it is grounded, evaluated and controlled.

• Architecture, user flow, data sources and access boundaries.
• Retrieval or tool integration with useful error and empty states.
• Evaluation evidence, known limitations and operational costs.
• LLM, embeddings, RAG, agent-design and project-based interview discussions.

Practical exercisePresent an assistant prototype with source-grounded responses, an evaluation report and a deployment plan.