Break a task into stages when each stage produces something you can inspect. Name its input, output and acceptance check. An ordered list makes work easier to review; the evidence supporting each result still needs checking.

Separate the task, then inspect the result

For a workshop announcement, first extract approved facts, then identify gaps, then write the message. Keeping these artifacts separate lets you catch an invented venue before it reaches the invitation. If the extraction fails, return to the source instead of polishing the draft.

This is an output workflow. You can request a source table, calculation or short explanation that a reader can inspect. You do not need the model’s private internal reasoning. OpenAI’s reasoning guidance recommends direct instructions for its reasoning models rather than requests to reveal a chain of thought.

Example input

“The community workshop is Saturday at 10:00 in Room 4. Bring a laptop. Booking is required. The booking link will be confirmed later.” The goal is a short invitation that stops before making up that link.

Use only the workshop notes below.
1. Extract the time, location, equipment and booking information.
2. List missing information needed for an invitation.
3. Write a draft under 60 words, leaving missing details explicitly pending.
Notes: The community workshop is Saturday at 10:00 in Room 4.
Bring a laptop. Booking is required.
The booking link will be confirmed later.

Example intermediate artifact

FieldSource valueCheck
WhenSaturday, 10:00Exact calendar date still needed
WhereRoom 4Matches notes
BringLaptopMatches notes
BookingRequired; link pendingAsk organizer for link

Example draft

Join the community workshop on Saturday at 10:00 in Room 4. Bring a laptop. Booking is required; the booking link is pending confirmation.

The draft preserves every supplied logistical fact. Before publishing it, obtain the exact date and booking link. A failed check is a reason to stop the workflow, not permission to infer those details.

When to make separate requests

One request is convenient for a small example. Use separate requests when a person must approve an intermediate result before the next stage begins. Anthropic’s prompt-chaining description uses successive calls with checks between them. You will build that handoff in the multi-step workflow lesson.

For practice, change the source so that it contains two conflicting times. A passing extraction reports the conflict and asks which is correct. It should not choose a time and carry it into the invitation.