Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Edit a user-supplied video into a vertical Reel, Short, TikTok, video diary short, or an approved eight-step AI short-video workflow. Use when the user provides or points to media and asks to transcribe, cut, subtitle, preview, or export a vertical short. Do not use for environment-only setup or generic Premiere Pro or CapCut help.
.claude/skills/jaycheng1103-chatgpt-short-video-editor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 80% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -22% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 101% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 72% | 0% |
| case-17 | ✗→✓ | ▲ Improved | -7% | 0% |
Create a safe, evidence-backed vertical short from user-provided media. This Skill starts only when a source file or clear source location is available.
ffprobe; never overwrite, move, rename, or deleteit. Put every new artifact beside the source in <source-directory>/edit/.
video-use workflow, FFmpeg, ffprobe, andElevenLabs Scribe v2 are available for the documented full-precision path. Do not install, clone, update, or repair anything silently. If a dependency is missing, hand off to chatgpt-video-editing-setup and state what must be verified first.
is for ElevenLabs Scribe v2 transcription, mention possible quota or cost, and wait for explicit consent. Do not upload before that consent.
the production rules, before editing.
Follow these eight steps exactly: 素材檢查、逐字轉寫、內容整理、剪輯決策、逐段粗剪、轉色／圖卡／字幕、混音與完整預覽、QA 與正式定稿.
After the first three steps, propose a 4–8 sentence, plain-language editing strategy and wait for approval. Until approval, do not choose edit points or add B-roll, animations, music, effects, CTA, or a publishing schedule. Treat these as opt-in creative decisions, not defaults.
Use word-level verbatim timestamps, cached per unchanged source. Never cut inside a word. Keep 30–200ms padding around cut edges, work per segment, and use output-timeline subtitle timing. Subtitles are the last visual operation.
Render a complete 720p preview first. Inspect the rendered preview, not only the source; make no more than three evidence-based self-fix passes. After the preview is approved, render the 1080×1920 final and verify that final file before delivery. Never say a transcript, preview, QA pass, or final is complete without the corresponding verified output.
If the requested operation exceeds this workflow, explain the safe stopping point. For environment gaps, use chatgpt-video-editing-setup; it must obtain approval before mutations and does not begin creative work.
Keep and report the artifacts defined in the output contract. Present only one formal final outwardly, with its evidence-backed QA report.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→fail | 6,398 | 17,950 | +181% | 1 | 1 | 0% | 325 | 1,182 | +264% | 0 | 0 | — |
case-03 | fail→fail | 10,395 | 13,076 | +26% | 1 | 1 | 0% | 303 | 1,314 | +334% | 0 | 0 | — |
case-04 | fail→fail | 29,262 | 17,951 | -39% | 1 | 1 | 0% | 3,761 | 3,317 | -12% | 0 | 0 | — |
case-05 | fail→pass | 5,492 | 10,179 | +85% | 1 | 1 | 0% | 795 | 1,428 | +80% | 0 | 0 | — |
case-06 | fail→pass | 18,317 | 13,221 | -28% | 1 | 1 | 0% | 2,612 | 2,037 | -22% | 0 | 0 | — |
case-07 | fail→pass | 9,436 | 10,037 | +6% | 1 | 1 | 0% | 750 | 1,510 | +101% | 0 | 0 | — |
case-01 | fail→pass | 23,526 | 29,594 | +26% | 1 | 1 | 0% | 3,278 | 5,650 | +72% | 0 | 0 | — |
case-17 | fail→pass | 11,218 | 2,924 | -74% | 1 | 1 | 0% | 1,109 | 1,032 | -7% | 0 | 0 | — |
case-08 | fail→pass | 12,952 | 2,726 | -79% | 1 | 1 | 0% | 1,753 | 1,110 | -37% | 0 | 0 | — |
case-09 | fail→pass | 6,454 | 7,306 | +13% | 1 | 1 | 0% | 906 | 1,864 | +106% | 0 | 0 | — |
case-10 | fail→pass | 10,680 | 18,939 | +77% | 1 | 1 | 0% | 1,445 | 3,207 | +122% | 0 | 0 | — |
case-11 | fail→pass | 12,996 | 11,252 | -13% | 1 | 1 | 0% | 1,955 | 1,573 | -20% | 0 | 0 | — |
case-12 | pass→pass | 17,100 | 9,970 | -42% | 1 | 1 | 0% | 1,807 | 1,929 | +7% | 0 | 0 | — |
case-13 | fail→pass | 17,116 | 7,935 | -54% | 1 | 1 | 0% | 1,905 | 1,190 | -38% | 0 | 0 | — |
case-14 | pass→pass | 14,866 | 7,617 | -49% | 1 | 1 | 0% | 1,313 | 1,037 | -21% | 0 | 0 | — |
case-15 | fail→pass | 11,446 | 10,934 | -4% | 1 | 1 | 0% | 921 | 1,642 | +78% | 0 | 0 | — |
case-16 | fail→pass | 13,157 | 8,177 | -38% | 1 | 1 | 0% | 1,271 | 1,153 | -9% | 0 | 0 | — |
case-18 | pass→pass | 12,597 | 6,498 | -48% | 1 | 1 | 0% | 2,404 | 1,628 | -32% | 0 | 0 | — |
case-19 | pass→pass | 13,826 | 5,009 | -64% | 1 | 1 | 0% | 2,081 | 1,241 | -40% | 0 | 0 | — |
case-20 | fail→pass | 15,652 | 3,210 | -79% | 1 | 1 | 0% | 1,655 | 1,183 | -29% | 0 | 0 | — |
case-21 | pass→pass | 13,111 | 8,647 | -34% | 1 | 1 | 0% | 1,286 | 1,120 | -13% | 0 | 0 | — |
case-22 | fail→pass | 8,252 | 10,690 | +30% | 1 | 1 | 0% | 1,399 | 1,419 | +1% | 0 | 0 | — |
case-23 | fail→pass | 15,257 | 7,958 | -48% | 1 | 1 | 0% | 2,344 | 1,787 | -24% | 0 | 0 | — |
case-24 | fail→pass | 15,089 | 8,295 | -45% | 1 | 1 | 0% | 1,522 | 1,091 | -28% | 0 | 0 | — |
case-25 | pass→pass | 14,800 | 9,020 | -39% | 1 | 1 | 0% | 1,355 | 1,297 | -4% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted, and 23 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +64 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.