Use a model to organize approved evidence, not to fill the gaps between sources with plausible-sounding facts. A clean draft should make uncertainty visible to its editor.
Separate source, inference, and unknown
Give the drafting workflow a bounded source pack. For each sentence with a factual claim, record one of three labels:
- Supported: a source directly confirms the statement.
- Inference: a conclusion drawn from several facts; show the reasoning and keep the language qualified.
- Unknown: no approved source supports it; ask for evidence or remove it.
A sample prompt rule might say: “If the source pack does not state the price, availability, result, or policy, write [VERIFY] and do not estimate it.” The reviewer should be able to trace every [VERIFY] item to an owner.
Example: describing a marketing method
A source says a GA4 report lists sessions by source/medium. That supports an instruction to inspect the report. It does not support a claim that “AI traffic increased 30x” or that the report proves which tool influenced a sale. Those require different data and a transparent method. Keep the explanation useful by showing the fields to inspect and the attribution limit.
For Arabic versions, preserve the distinction between an observed fact and an inference. Do not strengthen a cautious English phrase during transcreation. Check names, dates, currencies, and platform terms against the same source pack.
Review before release
Verify links, quotations, changing claims, and calculations. Ask a second reviewer to sample factual sentences and attempt to reproduce the conclusion. If the evidence is missing, update the source pack or publish a narrower answer; do not let the model invent a citation.
Optional AI-assisted work brief — Claim-to-source drafting workflow
Draft only from the source material supplied. Attach a source note to every factual, numerical, comparative, or policy claim. Mark gaps as questions for the owner instead of filling them from memory. Keep illustrative scenarios labeled as fictional. Return a claim ledger with wording, source, confidence, and reviewer action before producing polished copy.
Carousel draft: the claim ledger before the polished post
1 — Hook
A confident sentence is not evidence.
2 — Extract
List every number, comparison, guarantee, and policy claim.
3 — Attach
Record the source, date, exact support, and scope beside each claim.
4 — Resolve
Mark unsupported claims as questions for the owner; do not fill gaps from model memory.
5 — Draft
Write only after the evidence survives review.
6 — Close
Use AI for structure; keep evidence and approval with people.
Worked scenario — A claim-led drafting sequence
Start from a verified brief and create a two-column claim ledger: exact sentence and supporting source. The draft can be persuasive only inside the evidence boundary. For a hypothetical service, say “the team checks whether your area is covered” if that is the process; do not claim “coverage across Egypt” unless the coverage map proves it. Keep a reviewer question beside each unresolved claim.
Fact-check generated claims
Google advises reviewing AI-assisted content for accuracy, quality, and relevance before publishing. Keep the source ledger and human approval described above. Guidance on generative AI content.
AI disclosure: AI tools assisted with drafting. This example illustrates an editorial control, not a client finding.