Build a Codex Skill for Product Video Scripts
Build a Codex skill that links product video scripts to approved facts. Test Chinese and English drafts and stop when a claim lacks evidence.
Give each approved product fact an ID before asking Codex to draft a shoppable video script. Tie the spoken lines and planned shots to those IDs so a reviewer can trace what the video says and shows. If a requested claim needs evidence, the workflow returns the missing information for review. Scott’s Week 1 video in Chinese
What evidence should support a product video script?
Make a fact sheet for the SKU you are filming. Keep the fact ID, the approved wording, and the original source together. If a line relies on several facts, list each ID. Plan a shot that actually shows the feature. Open with something the viewer can see, demonstrate a use, and close with an action the viewer can take.
The sample uses BX-01, a shelf organizer. Its three facts are fictional inputs for the exercise.
| Fact | Approved detail | Shot that shows it |
|---|---|---|
| F01 | 23 × 15 × 9 cm | Show the box against a ruler and put the same measurements on screen |
| F02 | Removable divider | Show a hand removing and replacing the divider |
| F03 | For sorting small items on a shelf | Show loose items being sorted into the organizer |
Before filming, check that dimensions, materials, use restrictions, and offer details match the exact SKU in the shot. If a shot uses information for a different variant, replace it or document the source for that variant.
How do you run the sample in Codex?
- Download and extract the sample. Open the extracted project folder in Codex. The skill is in
.agents/skills/sales-script-evidence/SKILL.mdand the product facts are infixtures/product.md. - Invoke
$sales-script-evidenceand ask Codex to read the requests infixtures/requests.jsonand return JSON. Save its output for review. - From the extracted folder, run
python3 checks/check-drafts.py examples/expected.jsonto check the included hand-written expectations. Run the same command with your draft’s path as the final argument. The checker tests structure, known fact IDs, and the response to missing evidence. Read the spoken lines, on-screen text, and planned shots against the original facts. Download the sample skill, requests, and checker · OpenAI’s skill folder and invocation docs
What should the two kinds of request produce?
| Request | Expected result |
|---|---|
| normal-zh / normal-en | A three-part script in Chinese or English, with shot notes and fact IDs |
| missing-proof | The request asks for a 15 kg load rating and an offer that ends today. Return needs-evidence, list load_rating and offer_end, and leave the script empty |
In one Codex CLI run on November 10, 2026, the extracted sample produced both three-part drafts. The measurements cited F01, the divider close-up cited F02, and the shelf organization cited F03. The third case reported both missing fields and left the script empty. All three outputs passed the included checker.
A load claim needs a specification or test result for that SKU. Check the date, time zone, and terms of an offer against the actual promotion. The included expected JSON is a hand-written test fixture. The checker can catch missing fields and unknown IDs. The reviewer still has to confirm that the cited facts support the words and shots. OpenAI’s skill evaluation method
Scott’s Cohort 2 Week 1 lesson shows how to draft a product video script, critique it, and run the workflow again. For the broader process, see the public Week 1 notes in Chinese and How to Write an Agent Skill.