What to know before you read.
- A prompt is an interface specification, not a magic phrase.
- Reliable prompting requires examples, constraints, structured outputs and representative tests.
- Prompt quality must be measured against the task, not judged from one impressive response.
Prompt engineering is specification work
Prompt engineering is the practice of designing, testing and maintaining instructions and context so a generative model performs a defined task consistently enough for its intended use.
The strongest prompts make the job explicit: the objective, necessary context, boundaries, output structure and criteria for a useful result. They reduce ambiguity instead of decorating the instruction with dramatic language.
A repeatable six-step workflow
Begin with a task you can evaluate. ‘Write better’ is vague; ‘rewrite this support reply to include the resolution, next action and a calm tone under 120 words’ is testable.
- Define the task and user outcome.
- Collect representative ordinary and edge-case inputs.
- Write the minimum clear instruction and output contract.
- Add examples only where they resolve real ambiguity.
- Evaluate with a rubric and record failures.
- Revise one variable at a time and version the result.
Patterns that improve control
Use delimiters to separate instructions from source material. Ask for structured output when another system must consume the result. Require the model to identify missing information instead of inventing it. Include a refusal or escalation path for tasks outside scope.
Ground factual work in supplied sources and request citations that your application can verify. Telling a model to ‘be accurate’ is not a grounding strategy.
| Weak instruction | Stronger specification | Why it helps |
|---|---|---|
| Summarise this | Summarise for a project manager in five bullets: decision, owner, date, risk and next step | Defines audience and output |
| Extract the data | Return valid JSON with these fields; use null when absent | Creates a machine-readable contract |
| Answer accurately | Answer only from the supplied context and cite the source segment | Constrains evidence |
| Think carefully | Check the answer against this rubric before returning it | Connects quality to criteria |
Evaluate prompts like product behaviour
Create a small test set that represents the work: easy cases, ambiguous inputs, missing data, long context and adversarial wording. Score results using a rubric tied to usefulness, correctness, format and safety.
Track changes. A revision that fixes one example may damage another. Without a stable test set, prompt iteration becomes anecdotal and fragile.
Move from prompts to prompt systems
Once a prompt is reliable, learn retrieval, tool interfaces, model selection, latency, cost and monitoring. The prompt is one component in a larger system.
The Prompt Engineering Certificate of Specialisation develops this complete workflow through design, testing and maintenance. Generative AI is the wider application path; AI Agents is the next step when models begin using tools and state.
Questions readers often ask.
Does prompt engineering require coding?+
You can begin without coding. Testing at scale, structured outputs and application integration benefit from basic programming.
Are long prompts always better?+
No. Include only instructions, context and examples that improve the defined task. Unnecessary detail can create conflict and cost.
How do I know whether a prompt is good?+
Evaluate it on representative inputs using a rubric for correctness, usefulness, format and safety.
Roadmap