AI Skills Academy · Unit 8 of 8
Unit 8: Building with AI
Automating real tasks, structured output, tools and agents, and testing an AI workflow like a professional.
Opens the first lesson of this unit in AI Skills Academy. Every lesson is open from the start.
Lesson 1: Automate a real task
Goal: Map a repeatable task into an AI-assisted workflow with a human check.
The biggest everyday wins from AI come from repeatable tasks you do again and again, not one-off questions. Good candidates are tasks that take text in and put text out, follow a similar pattern each time and have results a person can check quickly.
- Turning messy notes into a to-do list
- Drafting replies to common customer questions
- Summarizing long documents into the same five headings
- Sorting feedback into categories
- Rewriting a weekly update for different audiences
Before you automate, map the workflow: input → AI step → human check → output. Write down what comes in, exactly what the AI should do, what the human checks and where the result goes.
Worked example: a small bakery's daily email
- Input: the day's customer emails, with names and contact details removed.
- AI step: "Sort these into orders, complaints and questions. For each, write a one-line summary and a suggested reply. Say 'unsure' if a message is unclear."
- Human check: the owner reads the summaries, fixes the suggested replies and handles complaints personally.
- Output: replies sent from the owner's own email, with each one checked by her first.
Twenty minutes of sorting becomes five minutes of checking, and every customer still gets a reply the owner approved.
Remember: Automate the repeatable middle. Keep a human check before anything reaches a real person.
Lesson 2: Ask for structured output
Goal: Get tables, checklists and JSON that other tools and people can use directly.
If the output will go into a spreadsheet, an app or a routine, ask for a structure instead of a paragraph. Name the fields and say what to do when something is missing.
Read this event flyer and return JSON with exactly these fields:
{
"event_name": text,
"date": "YYYY-MM-DD" or null,
"start_time": "HH:MM" (24-hour) or null,
"price_usd": number or null,
"location": text or null
}
Use null for anything the flyer does not say. Return only the JSON.
JSON is a simple text format for structured data that almost every programming language can read. Even if you never write code, tables and checklists work the same way: decide the columns first, then ask the AI to fill them.
Worked example: the output
{
"event_name": "Harvest Night Market",
"date": "2026-10-03",
"start_time": "17:30",
"price_usd": 0,
"location": "Old Mill Square"
}Every field has a known name and format, so it can go straight into a calendar or spreadsheet, after a quick check against the flyer.
Always validate. Check a few outputs by hand, and if a program reads the result, make the program check that every field is present and in the right format.
- Structured output
- AI output in a fixed format, such as a table, checklist or JSON, that people or programs can use directly.
- JSON
- JavaScript Object Notation: a simple text format for named fields and values that most software can read.
Remember: Name the fields, give the format and say what to use when something is missing.
Lesson 3: Tools and agents
Goal: Explain what tool-using AI can do and how to keep it safe.
On its own, a language model can only produce text. Many AI apps now connect models to tools: web search, a calculator or code runner, your files, your calendar, or a browser it can click through. An AI that plans steps and uses tools to get a goal done is often called an agent.
| Tool | What it fixes | New risk |
|---|---|---|
| Web search | Out-of-date knowledge | It may trust bad pages or hidden instructions |
| Code runner | Unreliable arithmetic | Code can still be wrong; check the logic |
| Your files and email | It can work with your real information | Privacy, and mistakes can reach real people |
| Actions (send, buy, book) | It can finish tasks, not just describe them | Mistakes cost real money or relationships |
The more an AI can do, the more the safety rules matter:
- Least access: give it only the tools and permissions the task needs.
- Approve actions: read what it is about to send, buy or delete before saying yes.
- Check sources: when it searches, look at the sources it used.
- Keep a record: know what it did, so mistakes can be traced and fixed.
Remember: Tools make AI more capable and more risky. Least access, approve actions, check sources.
Lesson 4: Test it like a pro
Goal: Evaluate an AI workflow on real examples before trusting it.
Professionals do not trust a prompt because it worked once. They test it on many real examples, including tricky ones, and measure how often it is right. This is called evaluation.
- Define success. What does a correct output look like? Write it down.
- Collect test cases. 10 to 20 real examples, including tricky ones: missing information, unusual wording, edge cases.
- Run and score. Compare each output to what correct looks like. Count the errors and note their types.
- Fix and re-test. Change the prompt, add examples or add a human check where errors cluster. Run the same tests again.
- Decide the human role. Which outputs can be used after a quick look, and which always need careful review?
Worked example: testing the bakery sorter
- Twenty past emails are labeled by the owner: 9 orders, 6 questions, 5 complaints.
- The AI sorts 18 of 20 correctly. Both mistakes are complaints that were written politely, which it labeled questions.
- The owner adds one polite complaint as an example in the prompt and re-tests: 20 of 20.
- She keeps reading every complaint herself anyway, because those matter most.
Testing turned "seems to work" into "right 20 out of 20 on our real emails, with a human check where it matters."
- Evaluation
- Testing an AI system on a set of real examples and measuring how often, and how, it gets things wrong.
- Edge case
- An unusual input that is likely to trip a system up.
Remember: It worked once is not a test. Run 10 to 20 real examples, count the errors and fix where they cluster.
Practise it in AI Skills Academy
Each lesson ends with two questions that check you can use the idea, not just recognise it. Wrong answers explain why and come back in Review. The unit ends with a 5-question quiz; pass it with 80% or more to master the unit.
Every lesson is free: no ads, no lives and no paywall.