The AI Toolkit an FDE Uses

Imagine you are the interviewer. You want to hire someone who can build with AI inside a customer's world. You would expect them to know the main building blocks and to choose between them with good reasons. This lesson is that toolkit. It covers prompts, structured output, search over documents, tools, workflows and agents, tests, and cost.

Start with the simplest thing

A strong candidate does not reach for the biggest tool. They start with the smallest design that could work. Then they add parts only when a test shows they are needed. Interviewers listen for this habit. It shows judgment.

The usual order is simple. Try a good prompt first. Add examples. Ask for structured output. Add search over the customer's documents if the model lacks the facts. Add tools if the system must act. Use an agent only if the path changes from case to case.

The building blocks, in plain words

• Prompt: the instructions you give the model. Clear instructions, a role, a format and a few examples go a long way.
• Structured output: you ask for a fixed shape, such as JSON with named fields. Your code can then check it and use it.
• Retrieval, or RAG: the system looks up the right documents first, then gives them to the model. The answer can cite its sources.
• Tools: functions the model can call, such as look up an order or create a ticket. Your code runs them and checks every input.
• Workflow: a fixed series of steps. It is easy to test, cheap and predictable.
• Agent: the model chooses the next step in a loop. It is flexible, and harder to test and control.
• Evaluation: a test set of real cases that you run after every change.
• Tracing: a record of every step, so you can find why a result went wrong.

Choosing an approach

Use a table like the one below to think out loud in an interview. The goal is to show the reasoning, not to recite a rule.

PROBLEM TYPE                         START WITH                          ADD ONLY IF NEEDED
Rewrite, summarize, classify         One prompt + examples              Structured output, a checker
Answer questions from documents      RAG with citations                 Better chunking, reranking, filters
Fill a form from messy text          Structured output + validation     A second pass for hard cases
Do a fixed multi-step job            Workflow of small steps            Retries, human approval
Handle many different requests       Router + a few narrow tools        An agent loop with a step limit
Anything risky (money, legal)        Human approval on the risky step   More tests, tighter limits

ALWAYS: a test set, logging of inputs and outputs, a limit on cost per task

What interviewers probe

• How do you stop the model from making things up? Good answers mention retrieval, citations, checks and an unsure path.
• How do you know it works? Good answers describe a test set, a pass mark and what happens when quality drops.
• What happens when it fails? Good answers include retries, fallbacks and a person in the loop.
• What does it cost? Good answers estimate tokens per task, then cost per month, and name ways to cut it.
• How do you keep customer data safe? Good answers cover least access, no secrets in prompts and permission checks in code.

Speed and cost basics

Long prompts and big models cost more and respond more slowly. You can cut cost in several ways. Send less text. Use a smaller model for easy steps. Cache repeated work. Run independent steps at the same time. Always measure before and after, so you know the change helped.