Prompt engineering for production AI
A prompt is a specification your business depends on. I write and audit prompts the way good engineers write specs: a clear role, boundaries, tools, output format and what to do when unsure, then I test them.
What a production prompt contains
- Role and goal
- Hard boundaries: what the agent must never do
- Rules for each tool it can use
- Output format that other software can rely on
- Tone and language, including Tamil and English
- What to do when it is unsure
- Worked examples of good and bad answers
An example
This is the shape of a real system prompt for a wedding-venue enquiry agent:
You are the enquiry agent for a wedding venue.
Answer only from venue_facts() and check_availability().
If you are not sure, say so and hand off to a person.
Never quote a price that is not in the price list.
Reply in the language the customer used (Tamil, English or mixed).
Collect: event date, guest count, budget, phone number.
Log every conversation to the CRM before you end it.What a prompt audit checks
- Ambiguous or conflicting instructions
- Missing refusals and missing hand-over rules
- Prompt injection risks, where a customer message tries to override the rules
- Tool misuse, such as calling the wrong tool or calling it too often
- Length and cost: tokens you pay for on every message
- Test coverage: which real cases have never been tried
Test sets and evals
Prompts are only as good as the tests behind them. I build a set of real scenarios from your business, including awkward and adversarial ones, and run every prompt change against it. You see what improved and what broke before anything goes live.
Deliverables
- Final system prompts, versioned and documented
- A test set and a before-and-after report
- Guardrail and hand-over rules
- Notes your team can use to change prompts safely
Frequently asked questions
What is prompt engineering?
Prompt engineering is writing and testing the instructions that tell an AI model how to behave. For production systems it means a precise specification: role, boundaries, tools, output format and fallbacks, tested against real cases.
I use ChatGPT or Claude already. Do I need a prompt engineer?
For personal use, no. For an agent that talks to your customers or acts on your systems, yes, because small wording differences change behaviour and a poor prompt can cost money or trust.
What is a prompt audit?
A prompt audit reviews an existing prompt for ambiguity, missing rules, security risks and cost, tests it on real cases and gives you a fixed version with a report.
How do you test prompts?
With a scenario set from your own business, run before and after each change. I include edge cases, mixed-language messages and attempts to trick the agent.
Keep reading
Tell me what eats your week.
My intake agent asks a few quick questions and puts your brief in front of me. I read every one.