ISTQB® CT-GenAI chapter 2: prompt engineering for effective software testing

In short

Chapter 2 of the ISTQB® CT-GenAI v1.1 syllabus is the heart of the certification. It covers how to build a prompt, the prompting techniques chosen for testing, how to apply them to each activity in the test process, and how to evaluate the results. At 365 minutes of training, it is by far the longest chapter in the syllabus.

It is also the most demanding one on exam day: it holds 5 of the 6 K3 questions (the sixth is in chapter 3), putting you in a test situation and asking you to pick or apply the right approach. Knowing the definitions is not enough; you need to have practiced.

What the exam expects

Chapter 2 accounts for 11 of the 40 questions, worth 16 of the 46 points: 6 K2 questions at 1 point and 5 K3 questions at 2 points. That is more than a third of the total score, with the pass mark at 30 out of 46.

Objective	Level	Questions	Points	What you need to be able to do
GenAI-2.1.1	K2	2	2	Recognize the six components of a structured prompt in examples
GenAI-2.1.2	K2	1	1	Tell prompt chaining, few-shot prompting and meta prompting apart
GenAI-2.1.3	K2	1	1	Distinguish system prompts from user prompts
GenAI-2.2.1	K3	1	2	Use generative AI for test analysis
GenAI-2.2.2	K3	1	2	Use generative AI for test design and test implementation
GenAI-2.2.3	K3	1	2	Use generative AI for automated regression testing
GenAI-2.2.4	K3	1	2	Use generative AI for test monitoring and test control
GenAI-2.2.5	K3	1	2	Select the right prompting technique for a given context
GenAI-2.3.1	K2	1	1	Understand the metrics used to evaluate results
GenAI-2.3.2	K2	1	1	Give examples of techniques for evaluating and refining a prompt

The five objectives in section 2.2 are worth 10 points between them. That is where the exam is won or lost.

Key concepts

The six components of a structured prompt. Role, context, instruction, input data, constraints, output format. Each answers a different question: who is speaking, in what situation, to do what, based on what, within which limits, in what shape. Sorting a prompt fragment into the right component is a classic question.

Three prompting techniques. Prompt chaining breaks a task into steps and checks each intermediate result before moving on. Few-shot prompting gives the model several examples so it reproduces a pattern (zero-shot means no example, one-shot means a single one). Meta prompting asks the LLM to generate or improve the prompt itself, which the tester then evaluates. The three can be combined.

System prompt vs user prompt. The system prompt sets the frame for the whole conversation (role, rules, tone) and is usually hidden from the user. The user prompt changes with every exchange. The model takes both into account together.

Generative AI across the test process. In test analysis: spotting defects in the test basis, deriving and prioritizing test conditions, analyzing coverage, suggesting test techniques. In test design and implementation: test cases, synthetic test data, automated test scripts, execution scheduling. In regression testing: keyword-driven scripts, impact analysis, self-healing tests, reporting. In test monitoring and control: metrics analysis, rescheduling support, test completion reports, dashboards.

Choosing a technique. A complex task where every step needs checking: prompt chaining. A repetitive task or an imposed output format: few-shot. A new task, or a prompt you do not yet know how to write: meta prompting.

Evaluating and refining. Seven metrics: accuracy, precision, recall, relevance and contextual fit, diversity, execution success rate, time efficiency. Five refinement techniques: iterative prompt modification, A/B testing of prompts, output analysis, integrating user feedback, adjusting prompt length and specificity. Because the LLM is non-deterministic, a single measurement proves nothing.

Common pitfalls

• Calling any long conversation prompt chaining. Ten messages in a row do not make a chain. What counts is the deliberate split into subtasks and the check on each intermediate output.
• Crediting meta prompting to the tester. If you rewrite your own prompt after a poor result, that is iterative prompt modification. Meta prompting means the model produces or improves the prompt.
• Picking the "most advanced" technique. In K3 questions, the right answer is the one that matches the task's characteristics (precision, repetition, novelty), not the one that sounds most sophisticated.
• Mixing up accuracy and precision. Accuracy measures overall correctness against a reference; precision measures correctness against a specific objective.

How to study it

Put the biggest share of your study time here, and make it practice. Take a user story from your project (anonymized) and write a prompt with all six components. Then process the same story three times: with prompt chaining (ambiguities, then test conditions, then a coverage check), with two few-shot examples, and by asking the model to propose the prompt for you. Compare the outputs using two or three metrics chosen in advance.

Start with the free level check to see where you stand. The certification guide places this chapter in the bigger picture, and the exam guide explains how the 2-point K3 questions are scored. The full course devotes several lessons to this chapter, with exercises for each test activity. For a deeper look at prompt structure, read our article on structured prompts for testers. Each technique also has its own hands-on article: prompt chaining, few-shot prompting for Gherkin and meta prompting, plus a complete method to generate test cases with AI.

Frequently asked questions

How many K3 questions are there in chapter 2?

Five, one per objective in section 2.2 (GenAI-2.2.1 to GenAI-2.2.5), each worth 2 points. They describe a concrete test situation and ask you to choose the best way to use generative AI, or the most suitable prompting technique. Together these five questions account for 10 of the chapter's 16 points.

Do I need to know prompting techniques beyond the three in the syllabus?

No. The syllabus singles out prompt chaining, few-shot prompting and meta prompting as the core techniques for testing, and explains the difference between zero-shot, one-shot and few-shot. Other techniques exist in the literature, but the exam focuses on these three and on how to choose between them for a given task.

Do the evaluation metrics involve calculations?

No, the objectives in section 2.3 are at K2 level. You are expected to understand what each metric measures and to recognize which one fits a described situation, for example the share of generated scripts that run without errors (execution success rate). There is no formula to apply.