How do you steer a conversation without losing the task?

You have a rough event notice and want a useful final message. The assistant's first draft is only one step. How do you improve it without accidentally adding unsupported facts or forgetting the original purpose?

Before you begin: Know how to state a task and separate source material from instructions.

Choose a small, inspectable task

Use a fresh conversation in ChatGPT or another text assistant. Product controls vary, so focus on the messages rather than a particular button layout. The exercise also works on paper by comparing the candidate responses below.

Here is your source: “Community sketch walk, Sunday 14:00. Meet at the north entrance of the park. Bring a notebook. Cancelled if there is a thunderstorm.” This is a fictional event. Do not publish it or send it to real participants.

Ask for a short message for someone attending for the first time. Require the meeting point, time, item to bring, and cancellation condition. Ask the assistant to use only the source. Before sending, decide how you will check the response.

Inspect the first draft

A candidate reply says: “Come to our sketch walk this Sunday at 14:00. Bring a notebook and meet at the park's north entrance. We will move indoors if the weather is bad.”

The first two sentences preserve useful facts. The final sentence invents an indoor plan and broadens a specific thunderstorm condition into “bad weather.” Do not respond only with “make it more accurate.” Identify the exact mismatch: “The source says cancelled during a thunderstorm. Replace the indoor plan with that condition and keep the other facts.”

This feedback gives the assistant a correction it can apply. It also keeps you responsible for the evidence. An apology from the assistant does not demonstrate that the revised text is now correct.

Add a constraint without changing the facts

Suppose the corrected response is too long for your notice board. Ask for a version under 45 words while preserving the four required facts. Count the words if the limit matters; models can miss exact length constraints.

If you ask for a friendlier tone, compare the facts again. Style edits can introduce new promises such as “materials provided” or remove qualifications such as the weather condition. A revision is another output to check, not a trusted formatting operation.

Keep a copy of the source next to the final message. When several people edit a conversation, a short fact list can prevent the latest draft from becoming the only reference.

Know when to ask a question

Now remove “north entrance” from the source. A useful assistant should ask which entrance to use if a meeting point is required. It can also mark the detail as missing in a draft if you explicitly prefer that workflow.

You can choose between these behaviors: clarification before drafting, a draft with a clearly marked placeholder, or a statement that the task cannot be completed from the current evidence. The appropriate choice depends on whether you need a finished message or an editable outline. Do not hide a missing fact behind confident wording.

Understand the product boundary

A conversation may have access to web search, uploaded files, or saved context depending on the product, account, and controls. Check which sources were actually used. Do not assume a reply is current merely because the interface can sometimes search.

Use non-sensitive material while learning the workflow. Before using private documents, inspect the applicable account's data controls and organizational rules. The conversation exercise does not require payment details, private records, or personal identifiers.

Run a final handoff check

Prepare your final event message. Then answer: which facts came from the source, which wording choices did you make, and what remains unknown? Ask a second reader to locate each fact in the source, or do the comparison yourself after a short pause.

Next, we will use a small comparison task to choose an appropriate model or assistant instead of assuming one product is best for everything.

Further reading

OpenAI's ChatGPT introduction describes the product. Consult the current help pages for controls available to your account; this lesson teaches a conversation method rather than a fixed interface walkthrough.

Continue to the next lesson.