AI from scratch for non-technical people

Create the account, decide which history you keep and leave a clean conversation to practise without dragging in old context.

Distinguish the technology that generates text from the company selling it and the interface you use. That order stops you confusing a bad answer with a bad model.

Move from a loose question to a three-part assignment: who you are, what you need and how you want to receive it.

Ask for specific facts with sources, check each one and note how the model behaves when it does not know. A confident tone is not evidence.

Classify your information as public, internal or sensitive, and write down which categories stay out of every tool.

Use a real assignment, try three assistants and score accuracy, format and correction effort. Choose by criteria, not habit.

Add a recipient, a goal, limits and a definition of "good" before asking for the text. Without an acceptance criterion there is no way to know whether it failed.

One example of what you want teaches more than ten adjectives. Prepare two good samples and one bad one to mark the boundary.

Define the columns or fields before generating. A fixed structure forces you to check data point by data point and makes the result easier to reuse.

Point to the exact sentence that fails and ask for a change only there. Rewriting everything from scratch erases what already worked.

Recognise generic text that meets the format and says nothing. It almost always lacks your context, not a longer instruction.

Consolidate role, context, task, format and criteria into a reusable template. Save it where you write every day, not in a lost note.

Upload minutes or a report and ask only for commitments, owners and dates. Check every line against the original before forwarding it.

Demand links, open each one and discard anything that does not appear in the source. A quote you cannot verify does not count as evidence.

Ask for a summary that keeps figures, conditions and exceptions, and flags what was left out for lack of space.

Order your notes, set the thesis and let the model propose a structure without inventing data you did not write.

Upload a CSV, ask for totals and anomalies, and check two figures by hand. AI speeds up reading, it does not replace control.

Pick a weekly task, turn it into a fixed assignment and measure how long you take before and after. If it saves no time, go back to the previous way.

Identify decisions with legal consequences, data you cannot expose and work where the process matters more than the result.

Compare plans, review what the provider does with your data and work out what happens if the price rises tomorrow or the account closes.

Choose one assistant, one search tool with sources and one document tool. Four badly used apps deliver less than two well integrated ones.

Write down which tool you use for each task, what you always verify and which data is off limits. Someone who was not in the course should understand it.

Decide whether your bottleneck is asking better, building tools or automating processes, and follow that.