Use Sol, Luna and Opus 5.5 to Build a Checkable Creative Brief

A useful AI creative brief connects each claim and shot to a source you can check. Sol, Luna and Opus 5.5 can be evaluated as text-workflow candidates, but handing a draft through three models does not itself make the claims accurate or produce a finished video.

This tutorial is for an individual creator turning a small source pack into a production plan. It uses a fictional desk-lamp brief, a 30-second shot list and a local validator. The source facts and example shots are authored teaching material, not outputs from a live three-model run.

We checked the Luna, Sol and Opus 5.5 prompting documentation on September 24, 2026. The division of work below is a proposed experiment, not a ranking or a claim that Ofox currently offers every model or tool.

Collect a source pack with explicit limits

Download the workflow lab and inspect creator-brief.json. It contains three sources. S1 says the fictional lamp has three brightness settings. S2 grants use of a supplied photo only within the fictional internal exercise. S3 requests a 30-second vertical storyboard and forbids price and eye-health promises.

In a real project, keep the original documents, dates, asset owner and allowed uses. A line saying an image is available does not establish advertising rights. An old product page may no longer support a price or feature. Put missing information into an unresolved list rather than inviting a model to fill it from memory.

Write down the audience, platform, desired action and constraints before generating copy. For example, a product explanation and a discount advertisement have different evidence needs. If the source pack has no verified offer, the brief must not invent a limited-time deal merely because it would make a stronger hook.

Assign a bounded role to each candidate

These assignments are starting hypotheses. A single model plus human review may be sufficient for your project. Test whether additional calls actually reduce correction time before making a three-stage system the default. None of these text roles replaces a licensed image source, an editor or a separate image/video generation service.

Give the reviewer the originals, not only the earlier model’s summary. Otherwise one omitted limitation can disappear through every stage. Keep the previous draft alongside revisions so you can see whether an accurate statement became an unsupported promise.

Use a prompt that preserves uncertainty

A reusable brief instruction is:

Use only the supplied source pack for factual claims.
For every product claim, include its source ID and supporting text.
Separate proposed creative choices from facts.
Put missing, contradictory or permission-limited items in unresolved.
Create a timed shot list matching the requested total duration.
Do not invent prices, endorsements, results or asset rights.

Do not paste private client data into an unapproved service. Use a source pack you are authorized to process. In the teaching example, all names and product details are fictional, so no real customer data or commercial permission is being demonstrated.

Check the shot list as a production artifact

The authored example has three shots: 0–6 seconds introduces a crowded desk, 6–18 seconds shows the three brightness settings, and 18–30 seconds places the lamp in the desk layout. Only the brightness line is a factual product claim; the scene choices are creative proposals.

Run the local check from the extracted directory:

python3 brief-check.py

It verifies that all referenced source IDs exist, the literal claim appears in a cited source and the shots form a continuous 30-second sequence. It does not verify that a photo was actually supplied, that the lamp exists, or that a voiceover fits the allocated seconds. Those are separate human and production checks.

Read the entire scene, including implied claims. A before/after sequence can suggest a health or performance benefit even when the caption avoids saying so. Inspect visual implications, subtitles, narration and the call to action together. For timing, read the voiceover aloud or measure an actual recording; a numerically valid shot list is only a plan.

Review claims and rights before generating assets

Use a ledger with claim, source, supporting passage, caveat, owner and approval state. Keep each unsupported item visible until it is removed or verified. A reviewer should be allowed to return no publishable claim when evidence is insufficient.

If an external tool creates an image or video, record that tool separately, with its settings and actual result. Do not label a generated scene as a real product demonstration, or infer the text model produced the media natively. For real product tutorials, capture the actual interface and record what you did; illustrative storyboards and real usage screenshots serve different purposes.

The JSON extraction guide provides a way to enforce the shape of the source index. The model selection guide explains why a candidate role still needs task-level evaluation.

Decide whether the workflow earned its extra steps

Before publishing, compare the draft with the original sources, clear the unresolved list or disclose remaining gaps, check links and confirm permissions. Measure unsupported claims, missing constraints, human correction minutes and total cost across all calls. Keep rejected drafts in the record.

For a small pilot, compare one-model drafting with the proposed organize–draft–review sequence using the same source pack. Judge the final artifacts against the same checklist. This article supplies the exercise and checks; it does not claim that more models improve quality, save money or guarantee a publishable result.