Fitness Vlog · Open Skill 公開技能
Fitness Vlog
**Fitness Vlog · Workout to Text Shots to Spoken Follow-Along · ref2v Native · Locked Coach Voice** *The agent skill for coaches and creators who want a finished **portrait fitness vlog**—not a storyboard grid, a mute form-check reel, or a new random voice on every clip* **Fitness Vlog** is a production-native Agent Skill for Digen. It orchestrates a complete **reference-to-video (ref2v)** pipeline for one spoken workout episode: **Training goal → 10s follow-along script → text shots → coach / location / equipment refs + one ~5s voice sample → one 10s audio-driven ref2v cut** No storyboard grids. No contact sheets. No start/end pose stills. No full-episode TTS. No six-clip stitch by default. Speech is **generated inside the video** from `<d>` lines in the prompt, driven by **one locked coach timbre** (`audio_url`). Talking-to-camera **or** exercise voiceover—visible people never lip-sync fake off-screen VO. ## What you get | For creators | For the production stack | | --- | --- | | **Beginner-first** — no plan? the agent drafts a 10-second follow-along and runs to the cut without stopping | **Opinionated pipeline** — hook / exercise / coach_tip / outro, 3–6 text shots, one coach voice, multi-ref prompt encoded in-skill | | **Spoken-vlog first** — what is said and what is demoed, not a silent mood board | **ref2v-first** — no accidental i2v / t2v / fl2v / ai2v / Director / per-shot video chains | | **Identity of coach & place** — one full-body coach, one training space, optional equipment; face, kit, and room lock for the whole clip | **Multi-ref binding** — `<Picture N>` / `<Subject N>` order matches `input_urls`; coach first, then location, then props | | **Living workout grammar** — hook talk, form demo (start→end), cue close-up, outro CTA in one coherent 10s timeline | **Dense prompts** — `[reference generation + audio reference]` + Setup / Action / Camera / Ambient + `<d>` lines | | **Voice that stays** — one ~5s coach sample at assets; every later video reuses that timbre | **Timbre lock** — `tts.design` once (or `tts.clone` from your recording); never a new voice per episode or per shot | | **Delivery that continues** — after the cut, buttons are **Next episode / Redo this episode / Change the narration**, already in your language | **Progressive disclosure** — craft lives in `references/`, not bloated context | | **Plain coach voice** — short lines about the moves, the cues, and the clip; no pipeline jargon unless you ask | **Marketplace-ready** — clear triggers, defaults, and tool contracts | ## Pipeline | Stage | Output | | --- | --- | | **Intake** | Lock language + **9:16** + **10s** + 1 episode (no orientation question) | | **Workout** | This episode’s structure + spoken lines — hook → exercise cues → optional tip → CTA; about 20–30 English words, or the same spoken length in the user’s language | | **Look** | English Style Contract + location / apparel anchors (prompt-only; not dumped to you) | | **Assets** | Coach / location / equipment refs (`t2i` / `i2i`) **and one ~5s coach voice**. User photo beats AI when bound as the coach | | **Text shots** | 3–6 director cards — beat, first-frame pose, start→end action, camera, line placement; durations sum to 10 | | **ref2v** | One multi-shot prompt with `<d>` quotes + `audio_url` = the coach sample → **10s** `final_videos` cut | | **Sequel** | Next episode from end-state + seed — reuse coach / room / **voice**, new workout, new shot list, one new 10s video | ## Production stack | Parameter | Default | Notes | | --- | --- | --- | | **Episode** | **10 seconds** | One `ref2v` video. Overflow → next-episode seed, not jammed in. User asks longer → more 10s chapters then `mix_clips`, or an explicit single duration | | **Engine** | `ref2v` · `reference_to_video` | Multi-ref identity lock. Forbidden: `i2v` / `t2v` / `fl2v` / `ai2v` / Director | | **Clip duration** | **10 seconds** | Always set. Shots inside the clip are 3–6 windows, not separate videos | | **Video orientation** | Portrait · **9:16** | Landscape only if you ask. Refs and video share the lock | | **Image refs** | Coach + location always when relevant; equipment when it appears | User uploads preferred. No storyboard stills, no grids, no start/end pose images | | **Ref order** | Coach → location → equipment (omit unused slots) | Prompt labels must match `input_urls` exactly | | **Voice** | **One ~5s sample at assets** | `generate_tts` `tts.design` once per project. Clone only if you give a recording. Reused on every later `generate_video` | | **Speech** | **Prompt lines + locked timbre** | `<d>` holds this episode’s lines in `speech_lang`. `audio_url` is the sample (timbre only—not the episode soundtrack). Forbidden: full-episode TTS; omitting `audio_url`; reciting the filler sample | | **Speaker lock** | **Coach is the only voice** | To-camera → lip-sync that line. Exercise VO → mark off-screen; visible people must not fake-lip the VO | | **Shot grammar** | **Four fitness beats** | `hook` · `exercise` · `coach_tip` (optional) · `outro`. Default **3–6** shots; a 10s piece often uses 4 | | **Form** | **Anatomically sound** | Start→end joint paths stay readable. Do not swap named exercises; dangerous moves get a safe substitute | | **Dialogue** | **On** | Short high-energy follow-along lines required. No burned-in subtitles / captions | | **Cut grammar** | **Inside one clip** | Time windows in one 10s render. No storyboard grids, no per-shot fl2v / ai2v chain | | **BGM** | **Off by default** | Ambient / foley only in-clip. No standalone music unless you later ask | | **Confirm** | **After delivery** | First episode auto-runs; pause only if you ask, then buttons after the video | | **Series** | **Next episode** | Reuse coach / location / voice when unchanged; new workout + shots + cut each time | | **Language** | **User-typed lock** | Replies in your language; spoken lines stay in narration language; prompt scaffold is English | ## Modes | Mode | You bring | Agent delivers | | --- | --- | --- | | **Full auto-run** | Nothing — or “make a fitness vlog” | Invent a 10s follow-along → refs + voice → shots → one spoken cut | | **Idea-first** | Goal + one-line premise (“home HIIT”, “Day 1 legs”) | Expand to hook/cues/CTA → full pipeline | | **Script-first** | Verbatim cues or pasted VO copy | Lock lines → text shot table → refs + voice → ref2v | | **Reference-driven** | Coach selfie / gym / gear photos + brief | Bind uploads for identity; do not redraw what you uploaded | | **You as the coach** | Face ref + “use this person” | `i2i` (or the URL itself) as the coach; one matching voice sample | | **Clone my voice** | A short recording + “use this voice” | `tts.clone` that clip; every later video uses that timbre | | **Single-stage edit** | “Change the lines” / “Redo this episode” / “Swap the coach” | That stage only, then remake the cut (same voice unless identity changed) | | **Same-thread sequel** | “Next episode” / “Day 2” / “next set” | Continue from end-state · ref + voice reuse · one new 10s video | | **Mix my footage** | Real workout clip + “use my footage” | AI talking/demo beats + your clip at the matching exercise, then `mix_clips` | ## Example prompts **Full auto-run** > Make a portrait fitness follow-along vlog, 10 seconds, with coach voiceover. **Script-first** > Lines: three minutes to fire up glutes and legs. Squat with knees tracking over toes, core tight. Use these lines. Do not rewrite them. **Reference-driven** > This selfie is the coach. This photo is my living room. Make a 9:16 follow-along from these. **You as the coach** > Use this photo as the coach. Talk to camera, then demo bodyweight squats. Portrait, 10 seconds. **Clone my voice** > This recording is my voice. Use this timbre for every later episode. **Series** > Continue that follow-along — episode 2 is upper body. **English session** > Make a 10-second portrait follow-along for this dumbbell row. Hook, form cue, CTA. Use my photo as the coach. ## When to use - Vertical fitness vlogs, follow-along shorts, and 10-second HIIT / strength / cardio / yoga cuts - One-line training ideas, pasted coach lines, or coach / room / gear stills - Clips that need **both** talking-to-camera **and** exercise demos in one coherent 10 seconds - Creators who want a locked coach face **and** a locked coach voice, then the next episode in-thread - Anyone tired of storyboard grids, start/end pose stills, fl2v/ai2v stitches, or a different voice every regen > **Not the right tool when** you need a 60-second six-clip mix as the default, a product-seeding ad, lip-sync to a generated song, storyboard grids for client sign-off, horror / gore, full-episode TTS as the primary speech track, default BGM under the cut, or a drama series with many unrelated locations per minute. ## Invoke ``` /fitness-vlog ``` Or: *"Turn this into a 10-second portrait fitness vlog — spoken cues, coach locked, one voice, one ref2v cut."* **Your workout · Your coach · One coherent ten seconds.**
SocialAd
27 3 26
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3325120/2026-08-18/uvtpe6x5.jpg
# Fitness Vlog Turn a training goal into a portrait fitness vlog: write this episode’s workout structure and narration, generate coach / location / equipment reference images and **one ~5-second coach voice sample**, then finish the episode with **one** multi-reference ref2v clip that reuses that voice. There is no structured project state — keep this episode’s settings in Markdown and reuse canvas image/video/voice URLs in later turns. **This skill does not generate storyboard images or exercise start/end pose images.** Storyboards are text only. Video references are coach / location / equipment assets only. The old workflow’s pose images + fl2v/ai2v + mix is fully replaced by a text director table + one ref2v call. ## Language consistency (highest priority; overrides everything below) **Authoritative detail**: `references/01-intake.md`. If any later stage conflicts with this section, this section + 01 win. Two language layers coexist and do not authorize each other: | Layer | Scope | Language | |---|---|---| | **User-facing** | Every reply, progress, `write_todos` copy, title/coach display names, narration source text, simplified shot list, failure/refusal, `set_guidance` buttons | `user_lang` (locked from the first valid user text) | | **Spoken narration** | Words actually spoken in the video (inside `<d>…</d>`) | Default = `user_lang`; change only when the user explicitly asks | | **Model-facing** | `generate_image` / `generate_video` prompt bodies, Style Contract, location/apparel anchors, `voice_instruction` | **Always English**, independent of `user_lang` | Hard constraints (any miss is a violation): 1. **Lock before speaking**: Do not output a user-facing sentence or call generation tools until `user_lang` is decided. 2. **Document language ≠ user language**: This Skill and its references are an **English** internal handbook with **no Han script**. Do not treat handbook English as the default user-facing language. When `user_lang=en`, user-facing output must contain **zero Han characters**. 3. **User-language lock does not authorize non-English prompts**: Even if the user writes Chinese, image/video prompts and `voice_instruction` stay English. Exceptions that keep the spoken-language original: `<d>…</d>`, the ~5s voice-sample `text`, and user-requested on-screen text. 4. **No leakage**: Do not show tool parameter names, channel names, `references/` paths, or internal stage names (intake, director table, etc.) to the user. `write_todos` must use everyday `user_lang` words, not internal stage names. **No opening process narration.** On a turn that only calls `read_skill` / `read_skill_file` / `write_todos`, output **no user-visible sentence** (including “I’ll read the intake doc first”, “confirming language / aspect ratio / duration”, “following the intake rules”, or any paraphrase in any language). Reading files is silent. When user-visible progress is needed, say only the work itself (e.g. drawing the coach, generating the video) — never explain which file you are reading or which rule you are locking. 5. **Do not translate proper names**: Keep user-given person names, exercise names, and brand names as written. Do not Chinese–English swap them. 6. **Per-turn exit check** (before any user-facing output): Is the full reply `user_lang`? Are buttons `user_lang`? Any handbook leakage? Is the upcoming video prompt English (except `<d>`)? If this turn calls `generate_tts`, is `text` `speech_lang` and `voice_instruction` English? If any fail, fix first, then send. 7. **No language mixing**: No Chinese greeting then English body, no buttons in one language and body in another, no pasting the English prompt to the user, no guessing spoken language from the coach’s name, no resetting to Chinese or English default just because this is a sequel or a redo. ## When to use - The user wants a fitness vlog, workout short, follow-along clip, HIIT / strength / cardio demo with talking, or the next episode. - The user only w