CyberFreezeDevExtraction
The same fields from every invoice or enquiry, with nulls instead of guesses.
Read more about Extraction
Getting the same fields from every invoice or enquiry, with nulls instead of guesses.
- Home
- AI glossary
- Prompt engineering
What is prompt engineering?
Treating a model's instructions like code: written deliberately, tested on real inputs, changed only on evidence.
The plain definition
A business prompt is a template of instructions, examples, retrieved data and user input. It is usually the cheapest lever, so push it before fine-tuning.
Read more about The plain definition
Prompt engineering is designing the input to a large language model so the output is correct, in the right format and consistent across thousands of calls. In a business system the prompt is rarely one sentence typed by a person. It is a template that combines fixed instructions, a few examples, retrieved data and the user's input.
It is usually the cheapest lever you have. Before considering fine-tuning, most tasks should be pushed as far as a well-built prompt will take them.
How a production prompt is built
-
01 Role and goal
Who the model acts as and what a good result looks like.
Read more about Role and goal
Who the model is acting as and what a good result looks like, in plain terms.
-
02 Rules and constraints
What to do when data is missing, what never to say, which language.
Read more about Rules and constraints
What to do when information is missing, what never to say, which language to reply in.
-
03 Examples
Two to five real input and output pairs show the format best.
Read more about Examples
Two to five real input and output pairs (few-shot prompting) show the format better than a description.
-
04 Context and input
Retrieved documents and record data, then the message, clearly labelled and separated.
Read more about Context and input
Retrieved documents, record data, then the actual message, clearly separated with labels or tags.
-
05 Output format
A JSON schema or exact template so downstream code can parse the answer.
Read more about Output format
A JSON schema or exact template, so code downstream can parse the answer.
Where it matters in business
Customer-facing replies
Tone, length and escalation rules for WhatsApp or email assistants.
Read more about Customer-facing replies
Tone, length and escalation rules for WhatsApp or email assistants.
Agents
An agent depends on clear tool descriptions and stopping rules in its prompt.
Read more about Agents
An AI agent depends on clear tool descriptions and stopping rules written into its prompt.
A practical example
A vague invoice prompt gave inconsistent results. The improved one names fields, sets formats, allows nulls and shows examples, then both run on 50 invoices.
Read more about A practical example
A first prompt says: "Extract the details from this invoice." Results vary: dates in three formats, GSTIN sometimes missing, amounts with and without commas. The improved prompt names each field, gives the date format (YYYY-MM-DD), says "if a field is not printed on the document, return null", shows two worked examples and asks for JSON that matches a schema. The team then runs both versions over 50 sample invoices and keeps the one with fewer errors.
Common mistakes
Editing without a test set
A fix for one case often breaks another. Compare on a fixed input set.
Read more about Editing without a test set
A change that fixes one case often breaks another. Keep a fixed set of inputs and compare. See model evaluation.
Contradictory instructions
Prompts grow by patching. Reread the whole thing and remove conflicting rules.
Read more about Contradictory instructions
Prompts grow by patching. Reread the whole thing and remove rules that fight each other.
Trusting the prompt for safety
"Never reveal X" is not access control. Enforce limits in code and guardrails.
Read more about Trusting the prompt for safety
"Never reveal X" is not access control. Users can talk models round. Enforce limits in code and with guardrails.
Questions
Does the same prompt work on every model?
Not reliably. Models differ in how they follow instructions and formats. When you switch model, rerun your test set and expect to adjust the prompt.
Is prompt engineering still needed with newer models?
Less trickery is needed, but clear instructions, examples and output formats still make a measurable difference, especially for structured business tasks.
Last updated