This skill turns one demanding video specification into a reusable Codex workflow. It activates only for narrated vertical shorts with both a final script and a final voiceover. It keeps the recording at original speed, routes the build through current HyperFrames guidance, stops at a voice-enabled Studio preview, and treats the encoded MP4 as a separate verification target.

What the ZIP contains

  • README.md gives project-scoped and user-scoped installation steps for macOS, Linux, and Windows.
  • SKILL.md defines the trigger, required inputs, production sequence, hard boundaries, and reporting contract.
  • references/production-contract.md carries the audio timing, vertical design, visual mapping, and project-record requirements.
  • references/quality-gates.md defines the preflight, strict browser check, Studio approval, and encoded-MP4 gates.
  • references/prompt-template.md points to the matching paste-ready production prompt.
  • assets/sample-30-second-voiceover.md supplies the original English Ten-Minute Reset script and fixed-title layout constraint.
  • assets/behavior-test-record.md records the retained build, validator, render, media facts, hash, and remaining evidence boundary.
  • agents/openai.yaml provides optional Codex interface metadata and a one-sentence starter request.

Why the final voiceover is required

The skill is deliberately narrower than a general video creator. It does not estimate scene timing from 79 words, assume a narrator speaks evenly, or build to a temporary synthetic read and swap the audio later. Missing final narration is a stop condition because every caption boundary, keyword response, transition, and end-card calculation depends on the real recording.

What remains flexible

The skill does not force the same palette, character, scene count, or transition on every video. It provides a restrained warm-gray vertical default, then follows stronger user direction. The reusable invariants are timing authority, factual preservation, mobile readability, semantic visuals, deterministic local dependencies, approval before render, and verification after render.

Installation boundary

Download the ZIP with the button above. Install it at project scope when only one repository should use the workflow, or at user scope when it should be available to every Codex project. The ZIP already contains the top-level codex-audio-first-short-video folder, so extract it directly into the selected skills directory.

Project-scoped install on macOS or Linux

Run from the project root
mkdir -p .agents/skills
unzip codex-audio-first-short-video-skill-v1-1.zip -d .agents/skills

User-scoped install on macOS or Linux

Make the skill available across projects
mkdir -p "$HOME/.agents/skills"
unzip codex-audio-first-short-video-skill-v1-1.zip -d "$HOME/.agents/skills"

Windows PowerShell

Project-scoped install
New-Item -ItemType Directory -Force .agents\skills | Out-Null
Expand-Archive .\codex-audio-first-short-video-skill-v1-1.zip -DestinationPath .agents\skills -Force

After extraction, confirm that .agents/skills/codex-audio-first-short-video/SKILL.md exists. Restart Codex if necessary, type /skills, and verify that codex-audio-first-short-video appears.

Start the skill explicitly

Starter request
Use $codex-audio-first-short-video to build a 30-second vertical video from my final script and narration audio.

Attach both the approved script and the exact final voiceover. The skill can also activate from its narrow description, but explicit $codex-audio-first-short-video invocation is easier to audit. The package includes the same installation instructions in README.md.

You can also download the retained sample MP4 before installing the workflow.

What was checked

Version 1.1 contains eight retained files and passes the bundled Skill Creator structural validator. A clean-install test extracts the ZIP into an empty .agents/skills directory, confirms the top-level folder and linked files, and validates the extracted copy. The workflow was also exercised on the Ten-Minute Reset: 28.500 seconds of natural-speed narration, six scenes, 37 captions, a strict HyperFrames check with no remaining errors or warnings, and an encoded 31.700-second H.264 MP4 with SHA-256 0d29144513ccd0160e9abe1433d21d3b24bd1fc8ae4c794e5c4421c8471ba12e.

What was not checked

The retained run used locally generated Kokoro narration and the authorized direct-render branch. It did not obtain independent human Studio approval, reproduce on a second host, upload to a social platform, or test high-stakes or rights-unclear source material. Its provider/ASR-derived timing is documented as such and is not represented as forced alignment. Those boundaries must remain visible when the skill is reused.