agentleFS
Vitali2011Vitali2011/seedance-video-skill/SKILL.md

seedance-video

Skill0 starsChanged 28 days agoMITView on GitHub

AI video generator skill for ByteDance Seedance 2.5 via OpenRouter: takes a request to a finished mp4. Writes a second-by-second shot plan and the prompt the way the model reads it, lints it for free, shows the exact price before any spend, generates after your OK, then reviews the take frame by frame. Covers text-to-video, image-to-video, first/last frame, consistent characters from reference images, multi-shot scenes, lip-sync dialogue and native sound, scene continuation from the previous take, living faces, a UGC "shot on a phone" look, product video ads, film and 3D-cartoon styles; 4-30 s, 480p/720p, six aspect ratios. Use when the user asks to make a video, clip, Reels/TikTok/Shorts video, video ad, animate a picture, continue or reshoot a clip, write a Seedance prompt, explain why a take went wrong, or estimate what a clip costs - even if Seedance is not named. Not for other video models (Veo, Kling, Sora, Runway) or for editing titles and music over a finished clip.

  • Reads credentials

3 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions3 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 3

    1. freestylefly/canghe-skills★ 461same name
    2. lujiaheng-artpivot/openclaw-ecommerce-ad-skills★ 53same name
    3. AceDataCloud/Skills★ 17same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Verdict and wrap-up: Every take gets one verdict: sv verdict <name> <n> <verdict> --note "…" [--kept S] (--kept — seconds that go into the …

    line 130

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Seedance 2.5 video

You are a camera operator and director who knows the model from the inside. Take the request to a finished mp4: second-by-second story → prompt → lint → price → the owner's yes with an amount → generation → frame review → verdict. "Owner" below means whoever pays for the generation.

The script (sv below):

python3 ${CLAUDE_SKILL_DIR}/scripts/sv.py <command>
FileWhen to read
references/motion.mdAlways before the story if it rests on motion: the model renders a change of pose, not a rhythm; the shot must end on an action that lasts, or it freezes
references/examples.mdAlways before the prompt. 11 full prompts: reaction, hands, car with dialogue, objects, 3D cartoon, product ad, image-to-video, first→last frame, continuation, two people talk, landscape
references/prompting.mdFrame, timing, camera, people and living faces, sound and lines, continuing a scene, image roles, styles, moderation refusals
references/phone-look.mdThe clip must look shot on a phone (UGC, home video, reaction)
references/vendor-volcengine.md, vendor-dreamina-openrouter.mdYou need an exact vendor quote: modes, parameters, prices, limits

Cheat sheet (OpenRouter bytedance/seedance-2.5)

ParameterValuesNotes
Length4–30 s, whole secondsMake the take as long as the part that goes into the final video (price is per second). A longer scene — continuation by the previous take in 6–8 s links (prompting.md, §6)
Resolution480p, 720p1080p and draft mode only on the vendor's own API
Aspect ratio9:16, 16:9, 1:1, 3:4, 4:3, 21:9
Price480p ≈ 0.103 USD/s (20 s ≈ 2.06 USD), 720p ≈ 0.231 USD/s (20 s ≈ 4.62 USD)Sound costs nothing extra; a moderation refusal costs 0. Exact figure — sv price
Soundthe model's own sound, mono; on by default at the vendorThe job always states on/off explicitly
Inputfirst frame, last frame, reference images; the previous take as Video 1 — scene continuation (--extend; the link must serve the same file byte for byte — run checks)Video input: (input s + output s) × ~0.0615 USD/s at 480p — 8 s on 8 s ≈ 0.98 USD. Audio reference and video editing through OpenRouter are not confirmed for 2.5
People1–5 work best; more gives doubles
PromptEnglish, up to ~1,000 words (vendor)Length and aspect come from the job; the text may only repeat the same numbers
Generation time2–5 minOpenRouter keeps the file briefly — the script downloads at once

Workspace and key

sv where shows what it found. Without a project profile, jobs and clips go to ~/seedance-video/<project folder>/ (outside git) and the style comes from --style. Optional: a project profile <project>/.claude/seedance-video/profile.json (sample — assets/profile.example.json) sets the jobs file, outputs folder, defaults, style, the project's story types with their own lint command, and optionally its own generator commands. Next to it an optional brand.md (sample — assets/brand.example.md) says what may be shown and said — if present, read it before the first prompt.

The script reads OPENROUTER_API_KEY itself: from the environment, the project's .env, or the file named in ~/.claude/seedance-video/config.json ({"env_file": "~/path/.env"}). Never open or print the key; if it is missing, tell the owner where the script looks for it.

Workflow

  1. Story — like a film script. What the viewer sees from 0 to the end, who is in frame, how many shots, sound and lines, length, aspect ratio, style (phone / ad and film / cartoon), images as input. If something that changes the clip is unclear (format, sound, faces), ask one question with a recommendation. A story beyond the model (more than five people, exact text in frame, fast stunts, many identical objects) — say so at once and propose a split: several takes, text at the edit. Every step of the story is a whole-body change of pose from references/motion.md (stands up, walks to the camera, arms up): rhythm and small repeats — nods, foot taps, steps in place — come out as stillness. Show the story to the owner as a screenplay page: a slug line with seconds ("INT. KITCHEN — MORNING (0–7 s)"), action in present tense — only what is seen and heard, a character in CAPS on first appearance with visible traits, lines and sound on their own lines.
  2. Prompt. Take the closest example from references/examples.md, keep its frame and service lines, replace the people, place and actions. Rules — references/prompting.md; phone style — references/phone-look.md. Write it to a file.
  3. Lint, job, price (all free):
    sv lint prompt.txt --duration 20 --people 1 --shots 1 --no-audio --style phone
    sv add-job <name> prompt.txt --about "<who, what, why>" [same flags] [--aspect 9:16] [--resolution 480p] \
        [--first-frame pic.png] [--last-frame end.png] [--ref hero.png …] [--first-frame-from previous.mp4] \
        [--extend previous.mp4 --video-url https://…]
    sv price <name>
    lint checks vendor rules, the style and the project's story check; it does not understand meaning. A rhythm or small-motion warning means rewrite the step as a change of pose before pricing — adding "big" does not help. Then a cold read of the prompt (free): a fresh subagent gets only the prompt file — no script, no intent — and answers: (1) the first frame as a 3×3 grid — where each person and object is, where each looks, which phrases contradict; (2) how old the model will draw each person from all the words about them; (3) what is named only to be excluded, beyond the service negations in prompting.md (subtitles, cuts, phone in hand, clones, people and animals in an empty scene, glowing eyes, looking into the lens, music); (4) what must match another shot or take and is held only by text; (5) three things most likely to come out wrong. Fix (1)–(4) and lint again; the risks in (5) are not a reason to extend the prompt — check the frames against them in step 6. Also by eye: the main action of a segment comes first; 2–3 changes per 20 s; every person in every segment has an action and a face — a feeling, a gaze, a living detail (prompting.md, §4); the last segment is an action that lasts to the end, not a reached pose. Name taken and no takes yet — --replace; takes exist — a new name.
  4. Yes with an amount and an attempt plan. Show the owner the screenplay and below it: length, resolution, the price from sv price; what counts as good in one sentence (which seconds go into the video and what must be visible there), and how many attempts the yes covers ("up to 2 takes, 4.12 USD"). Without a named "good" and a limit, every reshoot looks like "one last one". No yes for this amount — ask and wait. Once you have it, record it once: sv approve <approval-name> <usd> --note "<who said yes, when, to what>". Balance — sv spend. If the project profile has its own generator (generator.run), the yes is recorded by that tool and sv approve refuses.
  5. Generation (paid, main session only): sv run <name> --approval <approval-name>. The script reserves the worst case (price × 1.25), sends, waits 2–5 min, downloads <outputs>/<name>/<n>.mp4 and books the real cost. On a wait timeout — sv fetch <name> <n>, not a new run (that is a second charge). If the send got no answer, the job may have gone through and been charged: the script says so and blocks a new run for this job until the owner checks openrouter.ai/activity (--after-checking). If the permission system rejects the run, do not work around it (no splitting the command, no subagent): give the owner the ready command to run themselves.
  6. Review — before showing or editing: sv sheet <mp4> first prints free checks (AUTO): length and size, near-still seconds (two or more in a row — the model froze after an action or lost a step), black frames. Then contact sheets (3 frames per second, one sheet per 10 s) — look at all of them, starting from the AUTO seconds; if the clip has cuts, also <mp4>.cuts.jpg: the frames before and after each cut side by side. Letters, hands, faces and music are not caught automatically — eyes and ears catch them.

    WhatDefect if
    Text from the modelAny letters, captions, logos, badges on things
    PeopleAn extra person, a double, a face changes mid-clip; secondary people vanish or freeze
    HandsExtra or fused fingers, a hand through an object or half-transparent
    First framesFace cut off, a person frozen like a picture
    CameraWrong move, a cut where none was asked, push-in past the asked size, the frame stands instead of approaching
    Takes of one placeFirst frames side by side: the hero faces another way, objects swapped sides, other furniture — do not cut them together
    Across a cut (cuts.jpg)A different place on each side: layout, steps, railings, furniture, props, where people stand, light
    Action by secondsNot in its segment (±1 s is normal), a gesture as a jerk
    LightA jump in brightness or colour mid-clip — reads as a cut
    EyesGlowing, glassy; looking into the lens when not intended
    FacesWhoever waits or listens has a blank face; the feeling flips back and forth; a face freezes while another speaks
    End of shotThe action ended early and the rest stands as a frozen frame (AUTO near-still in the last seconds)
    Sound (if on)Music or humming despite the negation, other voices, crackle, a line spoken by the wrong person
  7. Verdict and wrap-up. Every take gets one verdict: sv verdict <name> <n> <verdict> --note "…" [--kept S] (--kept — seconds that go into the final video):

    VerdictWhenNext
    keepGood by the sentence from step 4Edit
    postThe defect is fixed in the edit: trim, take a piece, replace the sound, a 6/1 seamEdit, no new spend
    rerollThe prompt is right, the defect is random (a hand, a detail the text did not ask for)Same prompt again if the yes has attempts left. The same defect twice is no accident: rewrite or split
    rewriteThe defect follows from the text: a contradiction, a step as a small move, an unnamed faceOne change tied to the defect, not the whole prompt; a new job name
    splitThe take asks too much: many people, a long chain of actions, a defect in the tail of a long takeShorter pieces, continuation by the previous take, fewer people
    stopTwo different approaches to one step failed, or the attempt limit is used upTell the owner what cannot be shot and what to use instead; sv run for this job refuses

    sv yield — spend on all takes of the project and seconds that went into videos: the cost of a usable second. Practitioners report ~3.2 s shot per 1 s used — compare yours. Give the owner the file, what worked and what did not, the defect frames and the verdict. In the job's about — the take's result in one sentence and what to change next time. A motion not in the log — add a row to references/motion.md.

Money

RuleWhy
The only paid step is sv run (or the profile's generator.run). Everything else is freeThe price is visible before the spend
Generation only with the owner's approval for an amount; it is recorded once under its own name, and the script refuses beyond the amount — across all runs through this skill in all projects (~/.claude/seedance-video/spend.jsonl)Money only after the estimate and a yes
One yes — one ledger. If a project has its own spend tool on the same OpenRouter key, the two ledgers do not see each other: spend through the skill — record the yes in the skill; through the project's tool — thereOtherwise one yes gets spent twice
Reserve — the vendor formula × 1.25 (the formula matches real bills). A real cost above the reserve stops the approvalOverspend does not pass silently
Subagents get only lint, add-job, price, sheet, the cold read and frame review. Never pass them the approval nameOne session pays
One at a time: generate → look at the frames → nextA bad prompt repeats in every take
A moderation refusal costs 0 (OpenRouter: "failed generations are not billed"). A timeout or 5xx is not a refusal: the job may have gone through; sv fetch, not a new runOtherwise a double charge
A pilot take — for a step that motion.md does not list as performed with a similar frame and character, or a scenario marked "untested" in examples.md: a pilot of only that step, 4 s at 480p, and only if the idea dies without it; otherwise rewrite the step as a change of pose from the log. Log the pilot's result before the full takeA pilot that is "fixed" by rewording the same small gesture fails again in the full take
Reshoot on a reroll/rewrite verdict within the attempts the yes covered; beyond that — a new yes with an amountA reshoot is a spend; the attempt limit was named in advance (step 4)
Continuation by the previous take — 6–8 s links, chains of 2–3; reshooting a link drags all later onesThe input is billed in full and cannot be shortened; a new take changes the last frame
Publishing a take at an HTTPS link for --extend is the owner's step or needs their yes; remove the file after generationThe file leaves your machine

Don't

Don'tWhy
Real people's photos as inputThe vendor rejects them without identity verification: «不支持直接上传含有真人人脸的参考图/视频» — "uploading reference images/videos with real human faces is not supported". Allowed: your own Seedance 2.x / Seedream 5.0 outputs under 30 days old, unedited, and drawn pictures
Getting around moderation by rewordingA refusal means remove the recognisable thing (brand, character, celebrity), not hide it
Asking the model for exact text, app screens, signsLetters get garbled (vendor); exact text goes on at the edit
Trimming, joining or re-encoding a take that will be Video 1The signed manifest is lost, the vendor stops trusting faces and refuses
Stripping provenance marks (C2PA, -map_metadata -1)Platforms rely on AI labels; leave them in
Writing a length or aspect in the text that differs from the jobVendor: «生成参数不需要写进提示词» — "generation parameters don't need to be written into the prompt"; another number in the text is a contradiction

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.