The AI + SQL Workflow & Golden Rules

AI can sound confident and still be wrong. It can invent a column, use the wrong dialect, or quietly drop rows with the wrong JOIN. You are responsible for the final query. This topic is the habit that keeps AI useful and safe.

Definition

The AI + SQL workflow is a four-step loop: Ask a clear prompt → Review the answer → Test it on data → Refine with feedback. You repeat it until the query is proven correct.

The workflow: never copy-paste blindly

Step	What you do	Question to ask yourself
1. Ask	Write a clear prompt with context, task and specifications	Did I give the engine, tables and expected output?
2. Review	Read the answer line by line	Do the tables and columns actually exist? Is the JOIN type right?
3. Test	Run it on a small sample and compare with a result you know	Does the row count make sense? Are NULLs handled?
4. Refine	Found a problem? Paste the error or the wrong result back and ask again (prompt #5)	What exactly is wrong, and can I show the AI?

6 golden rules of AI + SQL

Rule	Meaning	Example
1. Be specific	Database engine, tables, columns, expected output	"SQL Server + players (id, name, team, country, runs)"
2. Give structure	Use "Do the following:" and a numbered list	1. 2. 3.
3. Ask why	Request an explanation so you learn, not just copy	"Explain each step"
4. Iterate	Refine the prompt with errors and feedback	See prompt #5
5. Always test	Verify on sample data and the execution plan	Never trust blindly
6. Stay safe	Keep sensitive data out of your prompts	Mask real data

How to test an AI query quickly

• Know one answer in advance. For example, Jos Buttler has the most runs (850). If the AI's "top scorer" query doesn't say Buttler, it's wrong.
• Check the row count. 6 players exist. A JOIN that returns 4 rows silently dropped 2.
• Look for NULLs. Does the query handle Warner's NULL team in player_profiles, or Dhoni's NULL innings?
• Check the dialect. LIMIT in a "SQL Server" answer means the AI ignored your context.

Data safety

Warning: Never paste passwords, connection strings, customers' personal data or confidential company data into a public AI tool. Share the structure (table and column names) and fake sample rows instead. That's enough for the AI to help.

Bonus: the 16 prompts at a glance

#	Prompt	#	Prompt
1	Solve an SQL Task	9	Improve Database DDL
2	Improve Readability	10	Generate Test Dataset
3	Optimize Performance	11	Create an SQL Course
4	Optimize Execution Plan	12	Understand a Concept
5	Debugging	13	Compare Concepts
6	Explain the Result	14	Practice SQL
7	Styling & Formatting	15	SQL Interview Prep
8	Documentation & Comments	16	SQL Exam Prep

Next up: real SQL projects. Put everything from this course together.

-- An AI says this returns "the top scorer". We KNOW it should be Jos Buttler (850).
SELECT TOP 1 name, runs
FROM players
ORDER BY runs DESC;

-- player_profiles has 6 players. Count the rows each JOIN keeps.
SELECT COUNT(*) AS inner_rows
FROM player_profiles AS p
INNER JOIN teams AS t ON p.team = t.team;   -- Warner (NULL team) is dropped

SELECT COUNT(*) AS left_rows
FROM player_profiles AS p
LEFT JOIN teams AS t ON p.team = t.team;    -- everyone kept

-- 🏏 Your turn: an AI wrote this to "list ALL players with their team name".
-- Run it. How many players are missing, and why? Change the JOIN to fix it.
SELECT p.name, t.team_name
FROM player_profiles AS p
INNER JOIN teams AS t
    ON p.team = t.team;