One-sentence positioning
The main router for the TikTok/UGC ad creative production pipeline. It does not replace any skill's independent invocation capability; it only helps you choose the right entry point based on the current task, or points out "what the next step usually is" after you get the output of a certain skill.
Mode one: pre-task routing (diagnose first, then point the way)
You say "I want to make a TikTok ad" → the router will diagnose what you currently have on hand and what is missing, then suggest which skill to start from. Each skill can be invoked independently, and there is no need to go in order—but if you do not know where to start, the router helps you locate it.
The following 8 skills have all passed independent verification and are registered in skills.json：
| Skill | Step | One-sentence description | Typical trigger scenario |
|---|---|---|---|
100x-video-reverse |
0 Video reverse-engineering | Local video → read-only evidence → semantic shot timeline + stable asset images + audio text evidence + prompt package + provenance | "reverse-engineer this video" "how do I replicate this video" "reverse this video" |
100x-segment |
1 Segmentation | Split a spoken script into three segments: paragraph logic + shot purpose + breath markers | "help me segment this" "where should I pause in this part" "break down this spoken script" |
100x-localize |
2c Ad copy localization | Source copy → Mexican Spanish localization, actively compressing to prevent translation expansion | "localize into Spanish" "translate into Spanish copy" "localize this ad copy" |
100x-persona |
3.1 Persona × scene | Script → "who tells it" + "where to shoot it," referring back to the original text word for word | "give this script a persona" "what scene is suitable for shooting this ad" "who should deliver this script" |
100x-exaggerate |
3.3 Exaggeration/contrast | Script → exaggeration techniques + contrast pairs, with intensity calibrated to the market/ceiling | "help add some exaggeration to this script" "how can it be more attention-grabbing" "add a before/after contrast" |
100x-search-query |
3.4 Search keywords | Product/persona → 15 English search terms each for Pinterest/TikTok/Reddit | "where to find reference images" "give me search keywords" "find reference images" |
100x-visual-fission |
5 Visual fission | Lock in person/scene/product references → media fission prompt matrix (single frame/first-last/first-middle-last/multi-day × ABC camera positions) | "help me fission this visual reference" "generate a fission prompt matrix" "fission this reference" |
100x-prompt-compose |
6 Prompt + model | Template + variables → final generation prompt rendered word for word and packaged for the target model | "help me assemble a prompt" "fill in a template to produce a prompt" "choose a template and configure variables" "compose a video prompt" |
Mode two: post-task navigation (you just finished X, what to do next)
You just finished running a skill and got the output → the router tells you the typical next step. Note: this is only a "recommended route," not a hard dependency—each skill can be invoked independently, and you can start from any entry point at any time.
| Just finished | Typical next step | Description |
|---|---|---|
100x-video-reverse |
→ Generate adaptation/100x-visual-fission／100x-prompt-compose |
First review the reverse-engineered package; the latter two Skills can consume the reference assets or prompts within it, but direct compatibility with the full reverse.json |
100x-segment |
→ 100x-persona or 100x-localize |
Segmentation is done—either pair with character scenes, or localize directly |
100x-localize |
→ 100x-exaggerate |
Spanish copy is ready—add exaggeration and contrast to keep it from feeling "flat" |
100x-persona |
→ 100x-search-query or 100x-exaggerate |
Character setting and scenes are set—you can search for visual references or design exaggeration points |
100x-exaggerate |
→ 100x-visual-fission or 100x-prompt-compose |
Exaggeration and contrast are designed—either do visual fission to produce multiple versions, or assemble the prompt directly |
100x-search-query |
→ 100x-visual-fission |
Reference images found—lock the identity and do medium fission |
100x-visual-fission |
→ 100x-prompt-compose |
Fission prompt matrix is ready—select a template + variables to assemble the final prompt and feed it to the model |
100x-prompt-compose |
(endpoint) | The prompt is assembled—you can feed it directly to Veo/Seedance/Jichuang |
Independent invocation guarantee
The above 8 skills each is fully independently invocable, with no requirement to run anything else first.100x-video-reverse
will internally run scripts/prepare_evidence.py(depends on ffmpeg/ffprobe) to establish a source fingerprint, dense frames and audio track,
then materialize and strictly validate the delivered media—this is its own workflow, not a requirement to run another skill first. Routing only helps you
choose an entry point and point the way; it is not a prerequisite. "Run routing first" is not mandatory—you can directly call any skill by name or say
the trigger word, and it will be activated automatically.
Skills that are not built / have not passed validation are not listed here
This routing lists only skills that actually exist and have passed independent validation. The remaining planned slots in the roadmap will be added here only after they are built.
Source
This file is the pipeline routing entry file. The skill list and trigger words are extracted from each skill's SKILL.md and metadata.json, all within this repository.