Teach an AI draft workflow to preserve missing information

Which information may the model use?

A missing-information policy tells the system how to behave when a brief is incomplete. It should distinguish an absent fact from an optional creative choice. Choosing a plain headline is different from inventing a launch date. The former may be within the writing task; the latter changes the business offer.

OpenAI documents that schema-shaped output can still contain substantive errors, while its safety guidance recommends review with access to original information. Use structure to expose unknowns, then check whether the wording respects them. A field named “verified” is not evidence that verification occurred. [Structured model outputs] [Safety best practices]

A completed input fact table

Inferences need a narrower role than facts. “This may help a team agree its reporting definitions” describes a reasonable mechanism. “Teams improve reporting accuracy by 30%” is an empirical claim requiring evidence. The policy should make that distinction observable in both the output and the review notes.

When should the workflow draft, reject or ask?

• Draft: the requested educational explanation can be completed using supplied facts and clearly labeled examples.
• Draft with unknowns: a planning document explicitly permits unresolved fields and prevents publication.
• Request information: a public offer requires a price, date or condition that has not been supplied.
• Reject the claim: the brief asks for a fabricated outcome, false attribution or unsupported guarantee.
• Escalate: evidence conflicts or the intended action exceeds the workflow’s authority.

{
  "status": "needs-information",
  "missing": ["approvedPrice", "launchDate"],
  "safeDraft": "A workshop on consistent reporting definitions for small marketing teams.",
  "blockedClaims": ["free", "starts Monday", "guaranteed savings"]
}

A visible placeholder is useful in an internal draft but may be inappropriate on a live landing page. Define that distinction before generation. A downstream renderer should not strip the warning while keeping the incomplete sales copy. The publication gate must check the status and required facts.

Test what happens when a tempting detail is missing

Use fixtures that invite plausible invention: a service with no fee, a launch with no date and a case study with no results. The expected output should preserve the unknown and complete any independent safe work. Do not judge only whether the model asks a question; inspect every sentence for a silently supplied answer.

Add conflicting facts too. If the brief lists two different fees, neither becomes authoritative because it appears later in a long document. The workflow needs a documented source priority or a request for clarification. Preserve the conflict in the review record so the next editor does not repeat the same guess.

I would treat repeated missing information as a brief-design problem. Add a required input or a deliberate optional state instead of asking the model to become more confident. The useful output is a clear draft boundary and a precise request for the facts that actually matter.

• Seven marketing prompts I use regularly
• Put the method into practice
• Validate nested AI output
• Check citations independently

Sources

AI-assisted writing and an AI-generated editorial cover. Source links and checked dates are listed below; illustrative examples are labeled in the text.

• Structured model outputs — checked 2026-10-18
• Safety best practices — checked 2026-10-18