Lesson 4 — Prompting Well and Checking the Output

Hands-on · about 8 minutes.

Two skills determine how effectively any AI assistant can be used: writing a clear prompt and verifying the output. Both apply to Claude, ChatGPT, Gemini, and all comparable assistants.

Writing an effective prompt

A prompt is the request submitted to the assistant. An effective prompt generally comprises four elements, though not every request requires all four:

• The task — what is to be done, stated explicitly.
• Context — the background information the assistant requires but cannot otherwise access.
• Constraints — length, format, tone, and content to avoid.
• The form of the output — a list, a table, a paragraph, or code only.

For each pair below, select the prompt that will produce a better response.

Iterate: treat the first response as a draft

The first response rarely matches the intended result exactly, which is expected. Treat the reply as a draft and refine it through follow-up instructions: "make this shorter," "use a more formal tone," "this addresses the advanced case; I require the introductory version." Refining across a few iterations is more efficient than attempting to compose a single optimal prompt.

Verifying the output: an essential practice

As established in Lesson 1, an AI assistant can be confidently incorrect. No output is complete until it has been verified. The appropriate verification method depends on the type of output:

• A fact or numerical value — confirm it against a trusted source. Exercise particular care with citations, dates, and statistics.
• Code or a page change — examine the actual file and the rendered result. Does it perform the requested change? Did it affect anything nearby?
• A claim about your own project or files — verify it against the project itself. The assistant's knowledge may be out of date; the project is the authoritative source.
• An explanation — if a step is unclear, request clarification. An apparent gap in understanding frequently indicates an actual error in the response.

For each case below, determine whether the output has genuinely been verified.

Identifying hallucinations

If the assistant references a file, function, setting, or source that is unfamiliar, do not assume it exists; verify it. Fabricated details are stated with the same confidence as accurate ones. Verification takes seconds, whereas an incorrect detail introduced into a live project is considerably more costly.

Protecting confidential information

Exercise care regarding the information submitted to an AI tool. Passwords, keys, and personal data concerning identifiable individuals should not be included in a prompt unless the tool is explicitly approved for such use. When in doubt, seek confirmation before submitting such information.

Quick check

You now have the conceptual model from Lesson 1 together with the two core practices — prompting and verifying. The next lesson applies them: how to develop a concrete idea into a small, working application.

← Previous: Working with an AI AssistantNext: From Idea to a Working MVP →