Lesson 3 — Working with an AI Assistant

Hands-on · about 8 minutes.

From conversation to collaboration

Lesson 1 addressed what an AI assistant is. This lesson addresses how to work with one on a concrete task — a document, a website, or a small program. Modern AI assistants can do more than answer questions: they can read files, modify them, and explain their contents. This capability makes them powerful and makes the manner in which they are directed consequential. The practices below apply to Claude, ChatGPT, and any other assistant.

1. Provide context before making a request

An AI assistant does not have access to your knowledge of the task. Before requesting a change, state what you are trying to accomplish and why. The objective is not a longer message but a message the assistant can act on without inference.

Weak Fix the page.

Better On my homepage, the "Contact us" button at the bottom links to the wrong page — it should go to contact.html. Update the link.

For each request below, determine which an AI assistant could act on directly and which would require it to infer the intent.

This activity needs JavaScript enabled. The lesson still covers everything.

2. Scope one task at a time

A request such as "redesign the entire site and add three features" is too large to execute well or to review effectively. Decompose the work into single, verifiable steps. Complete one step, review the result, then proceed to the next. Smaller steps are easier to verify and to reverse if the result is incorrect. This practice forms the basis of the build process in Lesson 5.

3. Exercise caution before high-risk actions

Reading files, drafting text, and editing a page are low-risk operations — straightforward to review and to reverse. Other actions are not: deleting files, modifying a database, or publishing content publicly. For any action that is difficult to reverse, proceed deliberately, confirm that it is the intended action, and seek assistance if uncertain. The principle is to verify before committing to an irreversible operation.

Classify the following actions: which may the AI execute directly, and which warrant review beforehand?

This activity needs JavaScript enabled. The lesson still covers everything.

4. Review the AI's changes rather than accepting them automatically

When an AI assistant makes a change, it provides a description of what it did. That description reflects the assistant's intent; the actual change is in the files. Open the file, or examine the result, and confirm that it matches the description. This practice detects the confident-error problem from Lesson 1 before it causes harm.

5. When a result is incorrect, describe the specific symptom

If a change is incorrect, describe the precise symptom — the expected behavior, the observed behavior, and any error message. This is substantially more useful than a general statement such as "it is broken," and it enables the assistant to identify the actual cause rather than infer it.

Quick check

This activity needs JavaScript enabled.

Summary: provide context in advance, work in small well-scoped steps, exercise caution before any action that is difficult to reverse, and always verify the actual result against the assistant's description of it.

← Previous: How AI Predicts the Next Word Next: Prompting Well and Checking the Output →