Steer with precise moves: museum caption

A vague starting prompt is:

Write a museum caption.

A plausible before output is:

This fascinating object changed history and remains important today.

The text sounds agreeable but cannot yet support a decision. It hides assumptions, supplies no inspection point, and does not show whether the result fits the real situation. Diagnose those defects before adding instructions. Prompt improvement is not decoration; each added phrase should control a known source of variation. This museum caption example is being used here to test format, examples, and questions.

For this page, use the following concrete revision:

Write a 90-word caption for visitors aged 10–14. Source facts: brass astrolabe, made in Cairo around 1300, used to estimate time from star positions. Define astrolabe; do not claim who owned it.

A more useful after output begins:

An astrolabe is a tool that uses the positions of stars to help estimate time. This brass example was made in Cairo around 1300. Its careful markings helped a user line up observations with a model of the sky. We do not know who owned this object.

The after output is easier to inspect because it follows explicit constraints and makes at least one uncertainty visible. Compare it with the before output line by line for museum caption: identify what came from source facts, what the model generated, and which decision still belongs to a person. Before acting, verify the claim with the highest consequence.

Judge the response against three criteria specific to museum caption: does it honor the requested form, does it rely only on supplied facts, and can the intended person act on it? Add a fourth criterion for format, examples, and questions. If a criterion matters, state a pass condition before asking for another draft so the model does not move the goalposts for you.

Remember the main limit: extra context can be stale, private, or distracting. A conversational response predicts suitable language from context; it does not inspect your home, understand institutional rules, call an expert, or accept responsibility. When the missing fact concerns safety, rights, health, money, assessment rules, or a relationship, turn the output into questions for an appropriate source. This museum caption example is being used here to test format, examples, and questions.

Suppose the first response invents one detail about museum caption. Quote the unsupported phrase and ask: “Keep the current format, remove that phrase, mark the missing fact as a question, and change nothing else.” This controlled follow-up tests format, examples, and questions while preserving material that already meets the quality bar.

Then ask the model to identify which statements came from your context and which it generated. Treat that labelling as an aid, not proof. Verify the highest-consequence statement using constraint coverage and unsupported assumptions. For the course case, write the source beside the checked statement and name who gives final approval. This creates a small audit trail that survives after the chat scrolls away. This museum caption example is being used here to test format, examples, and questions.

Minimise context before maximising it. Replace names with roles, remove addresses and account identifiers, summarise sensitive messages, and avoid uploading material you are not entitled to share. If the task can be completed with a blank template or offline checklist, that may be the better method. Relevance, not volume, is the standard. This museum caption example is being used here to test format, examples, and questions.

Set a stop rule for this museum caption exercise: stop after two targeted revisions if the response still invents constraints, ignores the format, or requires facts the tool cannot verify. At that point, complete the artifact yourself or consult a person. Knowing when conversation is no longer useful is part of proficient AI use.