female-outfit-director: a Codex Skill for multi-outfit collage first frames and beat-synced outfit-change videos

The gap female-outfit-director fills: prompt structure for outfit-change videos

Generating one good image of a person is a solved habit. Generating the same person five times in five outfits, arranging them into a single collage first frame, and then writing a video timeline where each outfit swap lands on a beat is where most people stall. The failure is not the model, it is the prompt. Identity drifts between outfits, the collage layout is described loosely, and the change moment has no defined mechanism, so the video model improvises a crossfade.

This Skill addresses that narrow gap. It is a Chinese-language workflow Skill for Codex or Skills-compatible agents, and its output is a production plan: parameter lock-in, a first-frame image prompt, a segmented video timeline, a condensed video prompt and negative constraints. The audience is prompt directors and small studios already paying for image and video generation and who want the text layer standardised rather than reinvented per project. It is not for someone who wants a rendered MP4 out of the box. The repository states plainly that it provides the prompt director workflow only and does not include third-party image or video generation services.

How the Skill structures identity, layout and the 12 change mechanisms

The workflow has three layers. The first is character locking: from a reference image it fixes facial features, hairstyle, temperament and identity consistency. When no reference image exists and no nationality is specified, the default is an original Chinese adult female character. That default is a design decision worth noticing, because it means an unspecified request never produces a neutral character, it produces a specific one.

The second layer is the collage first frame: five outfits for the same character, including the initial look and four changes, arranged as one central large figure plus four floating full-body figure cutouts. The default frame is 9:16, the default duration is 8 seconds, and the default camera is fixed. The third layer is the change mechanism. The README lists twelve options, including figure cutout fly-in, sleeve covering the lens, spinning relay, sticker page-turn and action beat.

The default mechanism, M1, is described as consumable cutout change with a copy-fly-in, consume-in-place rule. Clicking a target small cutout duplicates a full-body matte as a flying copy, the original position is emptied immediately, the copy moves toward the centre and scales up, aligns with the central figure at head, shoulders, waist and pose, triggers the change, and then both the flying copy and the white dashed outline disappear. The stated result is no leftover cutout and no ghosting on the central figure. That is a concrete rule set, and it is the most specific part of the README. The other eleven mechanisms are named but not spelled out in the README, so their depth has to be checked in skill/references.

Installing the Skill into Codex and running a first five-outfit request

The README gives a PowerShell install path. It clones the repository, creates the Skill directory under $HOME\.codex\skills\female-outfit-director, and copies the contents of the skill folder into it. The full steps are in docs/INSTALLATION.md.

git clone https://github.com/liyue-aigc/female-outfit-director.git
New-Item -ItemType Directory -Force "$HOME\.codex\skills\female-outfit-director"
Copy-Item -Recurse -Force ".\female-outfit-director\skill\*" "$HOME\.codex\skills\female-outfit-director\"

After copying, the README says to reopen the Codex task. The Skill is then invoked by name with a Chinese request. The README's own example asks for five urban outfits on the same Chinese adult woman, a collage first frame, and an 8-second beat-synced change video script.

$female-outfit-director 帮我为同一位中国成年女性设计五套都市穿搭拼贴首帧，并生成8秒卡点换装视频脚本。

What you should see is the output set the README describes: locked parameters, a first-frame image prompt, a segmented timeline, a condensed video prompt and negative prompts. The README points first-time users at docs/FIRST_USE.md, and that is where to look for the actual shape of each output block. User-specified parameters take priority over defaults, so the same Skill can be asked for a 15-second dance, a different number of characters, another camera or custom beat timings.

Where the defaults fight you: fixed camera, silent track, five outfits

The defaults are opinionated and they are the main limitation. A fixed camera across an 8-second video with five outfits leaves roughly 1.6 seconds per look, and the Skill's own mechanism list leans on fast actions like fly-in and page-turn to fill that. If your concept needs a moving camera or a longer hold per outfit, you have to override the default explicitly, and the README does not document what happens to the M1 alignment rule when the camera moves. The alignment step depends on the copy matching the central figure at head, shoulders, waist and pose, which is easier to guarantee in a locked frame.

Sound is the second constraint. The default is no background music, only the click, fly-in and beat sound effects the change needs. If you want a music track, the README says the arrangement follows your specification, which puts the beat-mapping work back on you. The third constraint is scope: the repository contains no image or video generation service. It is a text workflow, so every output still has to survive whichever model you paste it into, and the README does not claim otherwise. It is also Chinese-language, and the README does not describe an English mode.

Alternatives: writing the prompts yourself or using a general video tool

The realistic alternative is not another Skill of the same kind, it is doing this by hand. A prompt director writing five outfit descriptions, a collage layout and a change timeline from scratch gets total freedom and no inherited defaults. The difference is repeatability. female-outfit-director fixes the character-locking step, the five-outfit count, the 9:16 frame and the M1 consume-in-place rule so that two projects start from the same skeleton. Handwritten prompts start from whatever the writer remembers that day.

A general image or video generation product is the other alternative, and the difference is direction versus generation. Those tools own the pixels and usually the prompt format too; this repository owns neither. It produces structured prompts meant to be carried into a model you already use. If your model has a strong built-in character-consistency feature, the identity-locking layer here may be redundant, and the collage layout is the part you would still be buying.

Maintenance, licence and the cost of upgrading

The repository is not archived, and the last push was on 2026-09-07. There are no releases retrieved, so upgrades happen by pulling the branch and re-copying the skill folder, which is the same command the install uses. That is the upgrade cost: the copy step is destructive to the target folder, so any local edits inside $HOME\.codex\skills\female-outfit-director are overwritten by the Copy-Item -Recurse -Force call. The README does not document a rollback path or a versioned install, and CHANGELOG.md is the only place tracking what changed between pulls.

The licence is MIT, copyright 2026 liyue-aigc. MIT permits reuse and modification with the licence and copyright notice retained, and it comes with no warranty. For a prompt template that is a permissive fit. The licence does not cover the third-party image or video models you paste the prompts into, and the repository does not include those services, so their terms apply separately. This is a description of the licence text, not legal advice.

Editorial conclusion

Adopt it if you already run a Skills-capable Codex agent, work in Chinese, and want a repeatable prompt skeleton for one-woman, five-outfit, 9:16 change videos with a fixed camera. Skip it if you need an English pipeline, actual rendered video, or free camera movement, because the repository ships prompt structure and documentation only. Before committing, open skill/references to see how much of the twelve change mechanisms is actually written out, and check docs/FIRST_USE.md to confirm the output blocks match what your image and video models accept.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. liyue-aigc/female-outfit-director on GitHub
4. README