Pitch video auto generation

Product Pitch Video

Produce a finished ~30-second MP4: an Apple-launch-style product film whose visuals are generated shot-by-shot by Atlas Cloud video models, with an AI-generated instrumental soundtrack and crisp typography overlaid in post.

The pipeline (storyboard-first)

product URL ─▶ ① understand+derive ─▶ ② pitch script ─▶ ③ pitch.json
                (belief/contribution)   (beats+keyframes+motion)
   ─▶ ④ Phase A: keyframe stills ─▶ ⑤ HUMAN REVIEWS STORYBOARD ─▶ ⑥ Phase B: videos+stitch ─▶ final.mp4
       (chained image gen, ~$0.3)      (re-roll frames cheaply)       (parallel, ~$2.7)

You do the creative work in ①–③; scripts/pipeline.mjs does all the machinery (chained keyframe generation, parallel video generation, polling, download, overlay rendering, ffmpeg stitching). Never hand-roll API calls or ffmpeg commands — the scripts handle retries, resume, and cost reporting.

Why this shape:

• Shot-to-shot continuity comes from shared boundary keyframes: keyframe K_i is shot i's last frame AND shot i+1's first frame. Each keyframe is generated from the previous one in image-edit mode, so materials, lighting and the star object carry through the whole film. Videos are generated with first+last frame conditioning (image + last_image), and adjacent shots then cut seamlessly (hard cuts, no crossfade needed).
• Taste decisions happen at image prices. The storyboard gate (⑤) lets the user re-roll any composition for ~$0.04 before releasing the ~$2.7 video budget.
• AI video models cannot render text reliably, so prompts describe pure visuals (no text, no letters, no logos, no UI); all words go into overlays rendered by a headless browser at pixel-perfect quality.
• Everything is resumable: a crash or re-roll never re-bills completed work.

Classic mode (single video_prompt per shot, text-to-video, xfade transitions) still works for quick drafts — just omit keyframes from pitch.json.

Prerequisites (check before starting)

• ATLAS_CLOUD_API_KEY — env var, or .env file in the working directory (ATLAS_CLOUD_API_KEY=...). If missing, ask the user for their key. Never write the key into any file that could be shared/committed except .env.
• ffmpeg — on PATH, or set FFMPEG_PATH, or place a build at tools/ffmpeg/bin/ffmpeg.exe relative to the working directory. If absent on Windows, download the static build from https://www.gyan.dev/ffmpeg/builds/ffmpeg-release-essentials.zip and unzip to tools/ffmpeg/.
• Chrome or Edge — used headless for typography overlays (auto-detected).
• Cost awareness: tell the user the estimate before generating. Default 30s film = 6 shots × 5s at 1080p ≈ $2.7 video + ~$0.15 music (Seedance 2.0 ≈ $0.09/s). Testing/iteration? Use --res 720p and/or seedance-2.0-mini (half price) first. The pipeline prints actual billed tokens at the end.

Workflow

1. Understand and derive. Two scrapes, complementary:

   • Facts (text): WebFetch the URL for capabilities, numbers, price, tagline, tone of copy; grab the brand accent from the HTML (hex-frequency count works when compiled CSS has no semantic variables).
   • Temperament (visual): screenshot the rendered page and READ it — node <skill>/scripts/snap_page.mjs <url> --out page.png — then note dark/light feel, art direction (photographic? illustrated? abstract?), density vs whitespace, boldness. Text can't carry this, and it should steer the keyframe designs (an artful-dark brand tolerates bolder compositions than a clinical-light one).

   Then run the derivation protocol in references/pitch-script-guide.md: facts → what each removes for the user → the belief they point to → contribution sentences ("we did X, so you never Y"). The film argues the belief; it does not recite the feature list.

2. Write the pitch script. Read references/pitch-script-guide.md. Plan ~6 beats for 30s (belief → reveal → 2 contributions → testimony number → loop close). For each beat write the overlay copy (Apple voice, terse, full-width punctuation for Chinese) and the visual state the beat ends on — those end-states become the keyframes.

3. Design the keyframe chain + motion. Read references/shot-prompt-language.md, and pick the brand's archetype from references/brand-archetypes.md (dev-infra / palette-identity creative / hardware / luxury / fintech / wellness / entertainment / B2B SaaS) — it gives the world, texture words, palette policy, tempo and metaphor vocabulary as a starting recipe. N shots need N+1 keyframes where K_i = end of shot i = start of shot i+1, all featuring one continuous star object. K0's prompt carries the full style; later keyframes describe only the delta from the previous frame. Each shot gets a motion_prompt (pure motion + camera; frames own the composition). No text in any prompt.

4. Assemble pitch.json (schema below) in the working directory.

5. Run Phase A — storyboard.

   node <skill>/scripts/pipeline.mjs pitch.json --out pitch-build
   

   All keyframes are generated in ONE batch call to a sequential image model (bytedance/seedream-v5.0-lite/sequential) — cross-frame consistency is the model's job, and the whole set lands in a few minutes. The pipeline writes pitch-build/storyboard.html and stops for review. Show the storyboard to the user. To re-roll one frame: tweak its prompt, delete pitch-build/keyframes/<id>.png, rerun — single missing frames regenerate in edit mode with the previous frame as reference and the keyframe_style anchor appended, so re-rolls don't drift. A keyframe_style restraint anchor in pitch.json is mandatory — see the keyframe section of references/shot-prompt-language.md for the block and why image models need it.

6. Run Phase B — film. Rerun the same command once the storyboard is approved. Shots generate in parallel (first+last frame conditioned, 1–5 min each), overlays render, hard-cut stitch, soundtrack mixed → pitch-build/final.mp4. Resumable at every step; to redo one shot delete shots/<id>.mp4 + work/<id>.mp4 and rerun. Individual stages run alone (gen_image.mjs, gen_video.mjs, gen_music.mjs, overlay.mjs, stitch.mjs); references/atlas-api.md has the raw API for debugging.

7. Review and deliver. Extract a few frames with ffmpeg to sanity-check (text legibility, cut seamlessness at shared keyframes). Report the actual cost printed by the pipeline. Offer targeted re-rolls rather than full regeneration.

pitch.json schema (v2 — storyboard mode)

{
  "product": "Aura",
  "accent": "#7F72F7",
  "ratio": "16:9",
  "resolution": "1080p",
  "fps": 30,
  "cut": "hard",
  "model": "bytedance/seedance-2.0/image-to-video",
  "image_model": "bytedance/seedream-v5.0-lite/sequential",
  "edit_model": "bytedance/seedream-v5.0-pro",
  "keyframe_style": "<the restraint anchor — see shot-prompt-language.md>",
  "music": {
    "model": "minimax/music-2.6",
    "prompt": "minimal ambient electronic, evolving pads, subtle pulse, cinematic, luxurious, 90bpm, instrumental",
    "instrumental": true
  },
  "keyframes": [
    { "id": "K0", "prompt": "<full scene + style — text-to-image>" },
    { "id": "K1", "prompt": "The same scene and lighting. <only the delta> — edit from K0" }
  ],
  "shots": [
    {
      "id": "s1-belief",
      "duration": 5,
      "from": "K0",
      "to": "K1",
      "motion_prompt": "<pure motion + camera between the two frames>",
      "overlay": {
        "type": "statement",
        "text": "模态之间，{{不该有墙}}。",
        "in": 1.0,
        "out": 4.5
      }
    }
  ]
}

Classic mode: omit keyframes, give each shot a video_prompt instead of from/to/motion_prompt, and transitions become 0.5s crossfades ("crossfade": 0.5 to tune).

Overlay types (rendered by assets/overlay.html, same Apple type system as the launch-video skill): statement (one big line), hero (kicker + product name + subline), number (giant value + unit + label), closing (name + tagline + cta). in/out are seconds within the shot when the text fades in/out (leave ≥0.8s of clean footage at each end so crossfades never cut through text). {{braces}} in any text paint that phrase in the accent color. pos places the text block vertically: center (default), lower (lower third — use for statements over center-bright subjects), upper. Vary positions across the film and match them to the keyframe compositions (see the text-zone rule in references/shot-prompt-language.md). Top-level "overlay_theme": "dark" | "light" picks the type color: dark (default, white text) for dark films, light (ink text) when the brand's temperament calls for a bright film — white-on-light is invisible. Per-overlay theme overrides for mixed films.

Image model options: bytedance/seedream-v5.0-lite/sequential (batch mode — whole keyframe set in one call, best consistency) or a chain-mode family like openai/gpt-image-2 (no sequential variant; frames generate one-by-one, each edit-conditioned on the previous — slower but strong per-frame quality, and cheap). The pipeline picks the right path from the model name.

Quality bar

• Style block identical in every shot prompt; palette words match the brand accent. One motion per shot — no shot-within-shot montages.
• No text of any kind requested from the video model.
• Overlay copy survives the giant-screen test: short, declarative, specific.
• Chinese copy uses full-width punctuation (，。、) — half-width commas read as cheap.
• Total = shots×duration − crossfades; keep 28–32s for a "30 second" ask.
• Before calling it done, confirm final.mp4 exists, has audio, and its duration matches (ffprobe is printed by the pipeline).