TJ Short | Codex Ecommerce Short-Drama Skill

Use Codex to turn a product image into an ecommerce short-drama project: product analysis, target-region lock, character dossier boards, high-conflict scripts, 4x3 first-frame grids, 9:16 shot frames, model-specific prompts, captions, and delivery checks. The current default production route uses salpx gpt-image-2.5-sunburst for storyboard/first-frame images and salpx Wan 3.0 (wan3.0-*) for video clips.

中文版本

Overview

TJ Short is a public-safe Codex Skill template for ecommerce short dramas. It is not only a writing framework and not only a video API wrapper. It lets Codex move a product image through a complete production chain:

product analysis -> target-region lock -> three briefs -> product proof bible -> four libraries plus shot table (role / scene / product-prop / evidence + 12-shot production table) -> character dossier boards -> salpx gpt-image-2.5-sunburst 4x3 first-frame grid -> auto-cut 12 independent 9:16 first frames -> salpx Wan 3.0 prompts/API submission (text / first frame / first+last frame / multi-reference) -> captions and ad cutdown plan.

The split is clear:

• Codex is the production brain: judgment, writing, files, prompts, manifests, captions, reviews, and privacy checks.
• salpx is the image execution relay: runs gpt-image-2.5-sunburst for product-proof images, storyboard grids, and clean 9:16 first frames.
• salpx Wan 3.0 is the default and only video execution layer: takes accepted first frames or reference materials plus script-locked prompts and returns raw clips. Images and video share one SALPX_API_KEY.

Skill Card

Item	Description
One-line positioning	A Codex Skill for generating ecommerce short-drama projects, producing images through salpx, and producing video through salpx Wan 3.0
Input	Product image, product name, audience, product action, proof, CTA
Output	briefs, product proof bible, episode script, storyboard, first frames, video-model prompts, caption plan, ad cutdowns, US-market localization checklist
First-stage success	Validate 1 main episode plus 2 ad cutdowns before expanding into a series
Recommended video route	Default salpx / wan3.0-720p; optional wan3.0-1080p, wan3.0-hm-*, omni_flash, Seedance2, Veo
Not for	Pure hard ads, products without demonstrable action, projects without proof, or claims promising medical/financial results

It is useful for:

• pet products, consumer products, and tool products
• teams validating 1 main episode plus 2 ad cutdowns
• creators who want Codex to turn product assets into structured short-drama project files
• creators using salpx for storyboard/first-frame images and salpx Wan 3.0 for video (including 30s takes and image/video/audio references)
• teams selling to US audiences through TikTok Shop US, Amazon US, or other US ecommerce channels
• teams turning ecommerce short-drama production into a reusable skill or SOP

Core rule:

The product is not the hero. Characters, pets, relationships, and consequences come first. The product proves the truth later.

Version Highlights

Version	Key updates
v0.9.2	Banner replaced. The vector assets/salpx-banner.svg (which still described the retired MiniMax H3 route) is removed in favour of assets/salpx-banner.png (1280x427), generated end to end by gpt-image-2.5-sunburst with typography and imagery in a single pass. It now reflects the current route: salpx images plus Wan 3.0 video, 2-30s takes, native audio, lip sync.
v0.9.1	Default image model raised from gpt-image-2 to gpt-image-2.5-sunburst. .env.example, the image client default, and every SOP step were updated. API shape and flags are unchanged, so projects that pin the old model in .env keep working. Optional variants: gpt-image-2.5-flare, gpt-image-2.5-1k, and the XY- 1k/2k/4k lines.
v0.9.0	Video layer migration: MiniMax H3 is retired and salpx Wan 3.0 (wan3.0-480p/720p/1080p, plus the independent wan3.0-hm-* line) becomes the default and only video route. Images and video now share one SALPX_API_KEY. Added scripts/wan3_video_client.py (stdlib only; first frame / first+last frame / up to 10 reference images / 5 reference videos / 5 reference audios / document / web link; --dry-run quotes without spending) and docs/wan3-api-production.md. Video capability expands to 2-30s single takes, native audio, and reference-audio-driven lip-sync and singing. Added a hard cost gate: model, resolution, seconds, and price must be stated and approved before every submission. Deleted the H3 client and docs, dropped all H3_* variables, and made the client reject any base URL containing zhongbiao or h3.. Breaking for projects still calling the H3 route.
v0.8.29	Added a cinematic shot-design rule: seven elements (frame size, angle, composition, lighting, color grade, camera movement, pacing) and four reusable scene formulas (emotional build-up, suspense, romance, pressure) for the shot table's camera:/emotion_shift: fields; added a camera-as-witness prompt law (shooting process over finished result, exclusion lists, camera behavior on the timeline) as a selective realism layer that does not override existing Seedance2 motion/compliance rules.
v0.8.28	Added grid-first cost control: generate one 4x3 contact sheet first, then auto-cut 12 independent 9:16 first frames; do not generate 12 standalone first frames upfront unless explicitly requested or the grid/cut workflow fails.
v0.8.27	Corrected the Seedance2 visible-person first-frame flow: from the grid/first-frame stage, generate face_pencil strong / medium / blur_feature candidates for the same shot, precheck them one by one, then choose the passing frame with the most complete face and acting information; no-person product frames are only for API smoke tests or product-evidence shots.
v0.8.26	Added production-ready salpx API flow: gpt-image-2 image generation, omni_flash, seedance-2-mini-480p, and Veo video generation, plus a reusable client for submit, poll, and download.
v0.8.25	Added Volcengine Seedance2 prompt rules and model routing: Seedance2, omni, and Veo use separate prompt shapes, duration rules, references, and review strategies.
v0.8.24	Added Seedance2 filed role asset chain: character dossier board -> provider filing -> filed asset ID -> video generation with the filed asset; clarified that a watermark is not filing proof.
v0.8.23	Added character dossier board workflow: main portrait, three views, expression sheet, wardrobe/material details, product-contact details, and concise info panel as the standard role-subject example.
v0.8.22	Added AniShort-style subject libraries: role, scene, product/prop, and evidence libraries before the 12-shot production table.
v0.8.22	Added Seedance2 visible-face lessons: face_pencil and blur_feature before changing shot design.
v0.8.22	Added startup gate, salpx.com registration, and API Key setup reminders.

Full history: docs/changelog.md

Why Try It

• It is a Codex Skill, not just documentation
• Codex handles strategy, scripts, storyboards, first frames, prompts, and delivery checklists
• salpx executes gpt-image-2.5-sunburst image generation for storyboards and first frames; salpx Wan 3.0 generates video clips from accepted first frames or reference materials and prompts
• US-targeted projects are rebuilt for US characters, locations, dialogue, captions, CTAs, and platform context instead of translating Chinese ecommerce short-drama habits
• Avoids the weak "pain point -> product -> happy customer -> CTA" ad pattern
• Builds conflict, misunderstanding, and relationship pressure before product explanation
• Uses product as evidence: records, actions, procedures, behavior changes, or key objects
• Tests three clips first: hook, product evidence, ending hook
• Uses grid-first first-frame production: one 4x3 contact sheet -> auto-cut 12 independent 9:16 frames; standalone frames are for failed cells, selected retakes, or trial-clip polish only
• Treats Wan 3.0 as an integer 2-30s single-take route, defaulting to 5s for trials and 8-10s for story beats; omni remains an optional fixed 10-second route
• Separates video models: Wan 3.0 picks text_to_video / first_frame / first_last_frame / reference and labels materials as image 1, video 1, audio 1 inside the prompt; Seedance2 uses ordered media labels and subject definitions; omni describes motion after the first frame; Veo uses cinematic scene/camera language
• For Seedance2 visible faces, prefers character dossier board -> provider filing -> filed asset reference; a watermark is not filing proof
• Seedance2 visible-human projects must create dossier deliverables first: 角色主体库.md, 人物备案板需求.md, and Seedance2参考包计划.md
• Uses strict phases: startup only diagnoses and offers A/B/C; after selection, generate project files and dossier deliverables before storyboard grids, video prompts, or submissions
• Keeps character dossier boards for design, filing, and review; for video generation, extracts face close-up, full/half-body wardrobe, scene, product, motion, and audio references
• Seedance2 official first frames/grids should not downgrade the whole image to manga/anime. For each visible-person shot, generate face_pencil strong, face_pencil medium, and blur_feature candidates, precheck them one by one, then choose the passing frame with the most complete face and acting information.
• No-person product frames are only for API smoke tests or product-evidence shots; they must not silently replace visible-face acting shots.
• Seedance2 prompts must refer to media by order, such as 图片1, 视频1, and 音频1, rather than using asset IDs as character names
• If filing is unavailable or still rejected, uses face_pencil or blur_feature virtual-character repair before changing shot design
• Keeps pacing decisions in post-production
• Includes privacy and key-safety checks before public release

US-Market Localization

When the target region is US, United States, TikTok Shop US, or Amazon US, TJ Short locks the project to a US market context from the start. It does not translate a Chinese ecommerce short drama into English.

Required:

• Dialogue and subtitles default to natural American English unless the user explicitly requests bilingual output.
• Character names, jobs, household dynamics, conflict style, humor, shopping behavior, and CTA language must feel plausible for US short-form ecommerce.
• Scenes should use everyday US contexts such as apartments, suburban homes, kitchens, garages, dorms, pet stores, clinics, gyms, offices, driveways, and sidewalks.
• Phone UI, orders, pricing, time, address style, buttons, and CTA copy should follow US conventions and USD when pricing is shown.
• CTAs should use US-native wording such as Shop now, Check TikTok Shop, Get yours today, or Available on Amazon.
• Medical, supplement, financial, child, pet-health, beauty, safety, and other regulated categories need evidence-based phrasing and review notes where appropriate.

Blocked by default in US-locked projects:

Chinese livestream/ecommerce jargon
RMB pricing
WeChat/private-message CTAs
Chinese hospital, residential-compound, street-sign, or app-UI defaults
Chinese short-drama family/workplace tropes directly translated into English

If a US-locked artifact contains those mismatches, it is marked locale_mismatch_us_market and must be fixed before first-frame or video generation.

Seedance2 Visible Faces

Short-drama clips often need facial acting. Do not default to faceless crops.

Hard gate: create role dossier deliverables before first-frame grids and video prompts. A generic reference asset table is not enough.

Workflow gate:

1. Startup phase: product diagnosis and A/B/C only.
2. After brief selection: generate product proof bible, role subject library, character dossier spec, and Seedance2 reference package plan.
3. After the role gate passes: generate 12-shot storyboard, 4x3 first-frame grid, cut report, and model-specific prompt packs.

Cost gate: do not generate 12 standalone first frames upfront. Standard flow is 4x3 grid -> cut 12 frames -> precheck -> retake only failed cells or selected trial cells.

For Seedance2 visible-person official first frames, start the candidate workflow at the grid/first-frame stage:

1. If salpx or the selected provider supports role/person filing, file the character dossier board first, then use the provider-scoped filed asset ID for Seedance2 generation.
2. Keep the original filed platform reference; downloading and re-uploading may lose filing status.
3. Extract single-person references from the dossier board: 图片1 = face close-up, 图片2 = full/half-body wardrobe, plus scene, product, motion, or audio references.
4. Generate three candidates for the same shot: face_pencil_strong, face_pencil_medium, and blur_feature (blurred main frame + facial-feature sheet).
5. Precheck each candidate through Seedance2 or a controlled one-shot test.
6. Among passing candidates, choose the one with the most complete face information and clearest acting; do not resubmit a failed frame unchanged.
7. Define the subject in prompt text: 将图片1中的面部特征、图片2中的服装造型定义为林夏, then keep using the same role label.

Forbidden default: turning official first frames into full manga, anime, illustration, or commercial storyboard style when face review is risky. That is only a preview style board, not the production first frame.

No-person product frames are valid only for product-evidence shots or salpx API smoke tests. They are not the default production fallback for rejected visible-face acting shots.

Note: a LibTV-style watermark is not filing proof, and real filing IDs should not be committed to public repos. Public examples should use placeholders such as asset-YYYYMMDDHHMMSS-xxxxx.

Full SOP: docs/seedance2-face-compliance.md

How It Compares

Compared With	Where It Is Stronger	Where TJ Short Fits Better
OnlyShot	Deeper, heavier ecommerce short-drama production system	Lighter public Codex install for fast product validation
short-drama	Stronger for entertainment series and general drama structure	More focused on conversion, product proof, and ad cutdowns
Emily2040/seedance-2.0	Stronger video generation discipline and state tracking	Brings those ideas into a Codex ecommerce delivery chain
salpx video models	Execute omni, Seedance2, Veo, and other image-to-video generation	Codex handles judgment, scripts, prompts, manifests, and checks
Editing tools	Better for subtitles, dubbing, compositing, publishing	Better for creating the ecommerce short-drama structure from product assets

Detailed fair comparison: docs/comparison.md

Case Preview

Sanitized pet ecommerce short-drama example. These first frames represent the hook, product evidence, and ending hook.

Character Dossier Board

Key role subjects should not rely on a single attractive portrait. A reusable character dossier board locks the main portrait, front/side/back views, expression set, wardrobe/material details, product-contact behavior, and visual keywords. Later first-frame, storyboard, and video prompts should reference this role subject instead of reinventing the character in every shot.

Hook: send-away pressure	Product evidence	Ending hook

Example script: examples/xiderdl-lucky/ep01-high-conflict.md

Deliverables

Deliverable	Purpose	Required
Product proof bible	Defines user, product action, proof, and claim boundaries	Yes
Three briefs	Lets the team choose the story engine before writing scripts	Yes
Four libraries plus shot table	Role, scene, product/prop, evidence libraries, and a 12-shot production table	Yes
Character dossier boards	Lock key roles with main portrait, three views, expression sheet, wardrobe/materials, and product-contact behavior	Yes
Seedance2 filed asset table	Tracks fictional role filing status, filed asset ID placeholder, board version, and provider scope	Required for Seedance2 visible-face routes
High-conflict episode script	60-90 second episode with a strong first 5 seconds	Yes
12-shot production table	Each shot references subject libraries and focuses on action, emotion, camera, and dialogue	Yes
Three test first frames	Selects the hook, product-proof, and ending-hook frames from the 4x3 grid cuts for test generation	Yes
Image-to-video prompts	Script-locked prompts for salpx Wan 3.0 (declaring duration / ratio / material order)	Yes
Clip contracts	Defines what each clip can and cannot do	Yes
Reference role map	Separates first frame, product image, caption, and video reference duties	Yes
Generation manifest	Tracks model, frame, prompt, status, and output path	Yes
Voiceover and caption list	Source of truth for post-production subtitles	Yes
Caption plan	Ensures Wan 3.0 raw clips receive subtitles in post	Yes
Ad cutdown scripts	35-60 second paid-media cuts from the main episode	Recommended
Release scorecard	Decides publish, review-only, or rewrite	Recommended

Workflow Architecture

flowchart TD
  A["User in Codex: ecommerce short drama + product image"] --> B["Codex invokes TJ Short Skill"]
  B --> C["Codex product analysis"]
  C --> D["Codex generates three briefs"]
  D --> E["User chooses one story engine"]
  E --> F["Codex writes product proof bible"]
  F --> R["Codex builds four libraries: role / scene / product-prop / evidence"]
  R --> G["Codex writes high-conflict script + 12-shot production table"]
  G --> H["Codex creates role reference package + 4x3 first-frame grid"]
  H --> I["Auto-cuts 12 independent 9:16 first frames and prechecks them"]
  I --> J["Codex writes clip contracts + reference role map"]
  J --> K["Codex compiles salpx Wan 3.0 video prompts"]
  K --> L["salpx /v1/videos runs first-frame, first+last-frame, or multi-reference"]
  L --> M["Returns raw clips"]
  M --> N["Codex reviews story / product / lip sync / caption plan"]
  N --> O{"Did the three tests pass?"}
  O -- "Yes" --> P["Codex expands to all 12 shots"]
  O -- "No" --> Q["Codex one-variable retake or rewrite"]
  P --> R["Captions + dubbing + ad cutdowns"]

Environment

Item	Requirement
Codex	Install and run the TJ Short Skill
Git	Clone and version control
Python	Python 3.9+
Python dependency	requests
Image service	salpx relay / gpt-image-2.5-sunburst
Video service	salpx relay / Wan 3.0 (single route, shares the image key)
Recommended model	Default: salpx / wan3.0-720p; optional: wan3.0-1080p, wan3.0-hm-*, omni_flash, seedance-2-mini-480p, Veo variants
Aspect ratio	9:16
Clip duration	Wan 3.0: integer 2-30s, billed per second; default 5s for trials, 8-10s for story beats. omni optional fixed 10s; Seedance2/Veo follow model rules
Caption strategy	No burned-in subtitles during generation; add captions in post

Repository Contents

Path	Description
skill/SKILL.md	Skill file that can be copied into the Codex skills directory
README.md	Full Chinese documentation
README.en.md	Full English documentation
docs/comparison.md	Fair comparison with OnlyShot, short-drama, Emily2040/seedance-2.0, salpx, and editing tools
docs/changelog.md	Version changelog
docs/methodology.md	Ecommerce short-drama methodology
docs/privacy-and-release.md	Public-release privacy and key checklist
docs/seedance2-face-compliance.md	Seedance2 visible-face, face_pencil, and blur_feature compliance notes
docs/salpx-api-production.md	salpx API guide for gpt-image-2.5-sunburst storyboard grids, first frames, and product-proof images
docs/wan3-api-production.md	salpx Wan 3.0 guide: model and price ladder, billing and holds, material combination rules, submit, poll, download, error map
examples/xiderdl-lucky/	Sanitized sample script and screenshots
prompts/omni-fixed-10s-template.md	salpx / omni_flash fixed 10-second prompt template
scripts/salpx_production_client.py	Generic salpx image production client; salpx video remains optional fallback
scripts/wan3_video_client.py	salpx Wan 3.0 video submit, poll, and download client (stdlib only)

Installation

Option A: Install As A Codex Skill

git clone https://github.com/tttg2010/tj-short.git
mkdir -p ~/.codex/skills/tj-short
cp tj-short/skill/SKILL.md ~/.codex/skills/tj-short/SKILL.md

If you use WorkBuddy, copy skill/SKILL.md together with docs/, scripts/, prompts/, assets/, examples/, and .env.example into ~/.workbuddy/skills/tj-short/, then open a new session. No extra install command is needed.

After installing or updating the Skill, restart Codex. Otherwise Codex may still use an old cached Skill:

重新启动codex！
重新启动codex！
重新启动codex！

To fully test production from Codex, register at salpx.com and get an API key. Image generation (gpt-image-2.5-sunburst) and video generation (wan3.0-*) share that single key. Put it in local .env:

SALPX_API_KEY=your_salpx_api_key
SALPX_BASE_URL=https://www.salpx.com/v1
SALPX_IMAGE_MODEL=gpt-image-2.5-sunburst
SALPX_VIDEO_MODEL=wan3.0-720p

Without a salpx API key, Codex can still generate product analysis, scripts, storyboards, first-frame prompts, and manifests, but cannot submit image or video jobs.

Then say in Codex:

短剧带货启动

Or:

Use TJ Short to turn this product image into an ecommerce short-drama project.

Option B: Use The Scripts And Templates Only

git clone https://github.com/tttg2010/tj-short.git
cd tj-short
python3 -m pip install requests
cp .env.example .env

Register at salpx.com, get an API key, and fill in SALPX_API_KEY in your local .env — image and video share it. Follow .env.example. Keep real keys local only.

Never commit real keys. .env is ignored by git.

Usage SOP

Step 0: Start Inside Codex

Recommended starter:

短剧带货启动

After Codex gives A/B/C briefs, copy the next step:

短剧带货，选 A，地区 美国，图片模型 salpx / gpt-image-2.5-sunburst，视频模型 salpx / wan3.0-720p（可选 wan3.0-1080p、wan3.0-hm-*、omni_flash、seedance-2-mini-480p），产品是：your product in one sentence

Codex should analyze the product and produce three briefs first. It should not jump straight into a full script.

Startup gate:

• When the user says 短剧带货启动, Codex must not immediately generate a full production package.
• If no product image or product information is provided, Codex must ask for the product first. It must not invent generic workplace, fashion, or relationship examples.
• Even when the user says "generate video", follow this order first: product diagnosis -> three briefs -> user chooses A/B/C.
• Full production requires SALPX_API_KEY for both image generation and video generation (wan3.0-*).

Step 1: Codex Product Diagnosis

Answer:

• What are you selling?
• Who is it for?
• What product action can be shown?
• What proof makes it believable?

If product action and proof are unclear, do not generate video yet.

Step 2: Write Three Briefs

Each brief should include:

• story engine
• character relationship
• first 5-second crisis
• misunderstanding and truth
• product evidence position
• main selling point
• CTA
• AI video feasibility

Only after one brief is selected should you write the full episode.

Step 3: Codex Writes The High-Conflict Episode

Recommended rhythm:

0-5s: external pressure or relationship threat
5-20s: dialogue conflict
20-40s: misunderstanding escalates
40-55s: truth begins to appear
55-70s: product enters as evidence
70-90s: relationship turn + next hook

Step 4: Create One 4x3 First-Frame Grid, Then Cut 12 9:16 Frames

Do not generate 12 standalone first frames, and do not create three extra images solely for the trial clips. The standard cost-controlled workflow is:

1. Generate one cuttable 4x3 grid with 12 cells, one per episode shot.
2. Auto-cut 12 clean 9:16 first frames and save a cut report.
3. Precheck each cell for story, role, product, subtitle, and border issues.
4. Select HC-01 (opening hook), HC-09 (product evidence), and HC-12 (ending hook) from those cuts for the three trial clips.

Generate a standalone frame only if the grid cannot be cut, a cell fails precheck, the user explicitly requests it, or a selected trial clip needs a higher-quality retake. Official frames must be clean 9:16 full-frame images with no subtitles, inset, white border, or blurred background extension.

Step 5: Codex Compiles Prompts, salpx Generates Images, Wan 3.0 Generates Video

If you need to generate storyboards, first frames, or product-proof images first, use gpt-image-2.5-sunburst:

python3 scripts/salpx_production_client.py --env .env image \
  --model gpt-image-2.5-sunburst \
  --prompt "vertical 9:16 ecommerce short-drama first frame..." \
  --size 1024x1792 \
  --out-dir outputs/images \
  --name ep01-shot01

All video runs through salpx Wan 3.0 (wan3.0-*) using the same SALPX_API_KEY as image generation. The most common route is first-frame image-to-video: pass the accepted 9:16 first frame as first_frame.

python3 scripts/wan3_video_client.py --env .env video \
  --model wan3.0-720p \
  --prompt "8-second vertical 9:16 US TikTok Shop short-drama clip, no subtitles, no extra props..." \
  --duration 8 \
  --ratio 9:16 \
  --first-frame path/to/first-frame.png \
  --wait \
  --out outputs/videos/shot-01.mp4

Add --last-frame when the clip needs strict first/last-frame control (first_last_frame mode; cannot be mixed with reference materials):

python3 scripts/wan3_video_client.py --env .env video \
  --model wan3.0-720p \
  --prompt "Animate naturally from the first frame and finish on the last frame. No subtitles." \
  --duration 6 --ratio 9:16 \
  --first-frame path/to/first-frame.png \
  --last-frame path/to/last-frame.png \
  --wait --out outputs/videos/shot-01.mp4

Use reference mode for identity lock, camera transfer, voice, or lip-sync (up to 10 reference images; up to 5 reference videos totaling 15s; up to 5 reference audios totaling 15s):

python3 scripts/wan3_video_client.py --env .env video \
  --model wan3.0-1080p \
  --prompt "Image 1 is the lead, image 2 is the product pack; video 1 sets the camera move; audio 1 is her voice, lips must follow audio 1" \
  --duration 10 --ratio 9:16 \
  --ref-image path/to/character.png \
  --ref-image path/to/product.png \
  --ref-video path/to/dance.mp4 \
  --ref-audio path/to/voice.mp3 \
  --wait --out outputs/videos/shot-01.mp4

Pricing and the cost gate:

Model	Resolution	CNY per second	5s	10s	30s
wan3.0-480p	480P	0.21	1.05	2.10	6.30
wan3.0-720p	720P	0.42	2.10	4.20	12.60
wan3.0-1080p	1080P	0.84	4.20	8.40	25.20
wan3.0-hm-480p	480P	0.28	1.40	2.80	8.40
wan3.0-hm-720p	720P	0.56	2.80	5.60	16.80
wan3.0-hm-1080p	1080P	1.11	5.55	11.10	33.30

With a reference video attached, billing is price x (output seconds + reference video seconds). The submit-time hold assumes the 15s cap and is settled afterwards; failed or timed-out tasks are refunded in full. --dry-run validates materials and quotes without spending.

Before any submission, state the model, resolution, output seconds, whether a reference video is attached, and the computed price, then wait for consent. A 30s 1080p clip is CNY 25.20 and must never be submitted on assumption.

Full details: docs/wan3-api-production.md.

omni_flash (fixed 10s) plus Seedance2 and Veo remain as optional legacy routes. Their request shapes differ from Wan 3.0; do not mix them.

Step 6: Codex Reviews Clips

Check:

• Did the clip complete its narrative job?
• Did it reveal later information too early?
• Did the product become a hard ad?
• Did it create subtitles, garbled text, inset frames, or borders?
• Did dialogue belong to the right speaker?
• Does it need post dubbing?

Only expand to all 12 shots after the three tests pass.

Step 7: Codex Prepares Final Delivery

The final episode should include:

• subtitles
• dubbing or usable original audio
• low-mixed original sound
• preview frames
• release score
• ad cutdowns

Raw clips without subtitles are not final publishable videos.

Star Rating Criteria

If a user can complete these actions within 30 minutes, the skill deserves 5 stars:

Rating	Standard
1/5	Concept only; unclear how to start inside Codex
2/5	Codex can write a script but cannot enter video production
3/5	Codex can create first frames and prompts but lacks review criteria
4/5	Codex can run three test clips and knows how to retake failures
5/5	Codex can go from product image to main episode, cutdowns, captions, and release checks

References

This workflow openly credits:

• OnlyShot: ecommerce short-drama thinking, product evidence, projectized delivery
• short-drama: episode structure, storyboard, video-production workflow
• Emily2040/seedance-2.0: clip contracts, project state capsule, reference role map, one-variable retake

This is not an official distribution of those projects. It is a public-safe reusable practice template.

Keywords

Codex skill, Codex ecommerce short drama, ecommerce short-drama skill, shoppable short video, AI product video, product image to video, image-to-video, salpx Wan 3.0, wan3.0, 30-second video, multi-reference, reference video, reference audio, lip sync, salpx, salpx relay, gpt-image-2.5-sunburst, omni_flash, Seedance 2.0, seedance-2-mini-480p, Veo, AI video generation, short-drama storyboard, 4x3 first-frame grid, 9:16 first frame, character dossier board, role asset filing, Seedance2 face compliance, face_pencil, blur_feature, US-market localization, product proof, ecommerce video workflow, short-form commerce, TikTok Shop content, cross-border ecommerce video, prompt engineering.

Public Safety

This repository does not include:

• real API keys
• .env
• private product source images
• API responses
• task IDs
• download URLs
• local absolute paths
• generated video files

Read: docs/privacy-and-release.md