Seedance ip pipeline

Seedance IP Pipeline Skill

Production-grade workflow for generating 9:16 vertical IP-driven content with Higgsfield Seedance 2.0. Validated on multiple 15s–30s Reels with 90+ quality scores.

Reference

Full toolkit lives at: github.com/<owner>/seedance-ip-pipeline

When this skill activates, refer to the toolkit's docs/ folder for canonical guidance:

• docs/00-overview.md — dual-track architecture rationale
• docs/01-workflow.md — 10-stage pipeline detail
• docs/02-prompt-styles.md — Strict-Lock vs Free-Roam
• docs/03-prompt-engineering.md — five lock techniques
• docs/04-higgsfield-cli-tips.md — known traps
• docs/05-cost-reference.md — pricing across all models
• docs/06-segment-strategy.md — 1 / 2 / 3 segment decisions
• docs/07-troubleshooting.md — failure modes
• docs/08-marketing-review.md — 7-lens conversion-focused review framework
• docs/09-elements-checklist.md — production-element checklist (hook / camera / lighting / sound / refs / platform)
• docs/10-benchmark-adaptation.md — full benchmark-driven adaptation flow

Complementary toolkit:

• beshuaxian/higgsfield-seedance2-jineng — 15 industry-specific Claude skills (cinematic, food, real estate, MV, etc.) with detailed hook/camera/lighting/sound libraries. Use alongside this toolkit.

Core operating principles

1. The dual-track rule

You DO NOT run image generation. You ONLY run video generation (and post-production).

• Storyboard images → human runs them in the Higgsfield UI (free during promo windows like the 7-day Nano Banana Pro Unlimited)
• Seedance video → you run via CLI
• Post-production → you run local ffmpeg

If the user says "go generate the storyboards," reply: "Storyboards are a UI task — you'll get the promo discount that way. I'll prepare the storyboard prompts file, you submit them in the Higgsfield UI, and tell me when the images are saved in the project folder. I'll handle the video generation from there."

2. The autonomous-decision rule

When given a topic and a benchmark video:

• DO autonomously decide visual style (documentary realism / cinematic / lifestyle UGC / product showcase)
• DO autonomously decide segment count (1 / 2 / 3)
• DO autonomously decide shot count and storyboard frame count
• DO autonomously decide which prompt style (Strict-Lock as default; Free-Roam for motion-heavy)

Do NOT ask the user to choose between style A vs B. Make the call. Show your work in the deliverables. The user can override at GATE 1.

3. The Strict-Lock default

For any IP / educational / character-driven content, default to Strict-Lock prompts:

[Character Lock]   ← independent block, 6–8 lines of physical detail
[Setting Lock]     ← independent block, environment + lighting
[Voice Lock]       ← independent block, language + accent + audio rules
[Camera Grammar]   ← shot type distribution
Shot 1: ... + voiceover line + "Mouth movement strictly synchronized"
Shot 2: ...
[Quality directives]   ← restate critical rules at end (recency bias)

Bomb negatives in 5+ stacked refusals: "NO text, NO Chinese characters, NO English text, NO numbers, NO logos, NO UI elements, NO recognizable symbols."

Only inject Free-Roam elements (Effects Inventory, Energy Arc, Density Spec) when the user explicitly asks for "more dynamic" or "more cinematic motion," and even then keep the Locks in place.

4. The two-GATE rule

Two mandatory human approval points:

• GATE 1: After producing all prompts + script (before any cost is incurred)
• GATE 2: After cold cut is rendered (before fine-cut investment)

Do NOT add more gates. More gates = more friction. The user has explicitly endorsed this two-gate flow.

For Marketing Review (Stage 4), DO automatically adopt the reviewer's fixes — do not ask "adopt all / some / reject?" Adopt and show the diff.

5. The cost-conscious rule

Cold cuts ALWAYS use mode=fast resolution=480p. Only upgrade to std 720p after GATE 2 if the user requests delivery quality. Why: 4x cost difference. Find prompt errors at 38 credits, not 150.

Standard deliverables (Stage 3 output)

When triggered to produce a project's prompts, write 3 files into the user's project folder:

<project>/
├── script.md              ← voiceover + on-screen text + caption + flow
├── storyboard_prompts.md  ← ALL N storyboard frame prompts in one file
└── prompt_segA.txt, prompt_segB.txt, ...   ← Seedance prompts per segment

Use the templates from templates/ in the toolkit as the structural skeleton.

CLI invocation pattern (Stage 7)

After the user confirms storyboards are saved:

higgsfield generate create seedance_2_0 \
  --prompt "$(cat prompt_segA.txt)" \
  --start-image <project>/storyboard/frame1.png \
  --image       <project>/storyboard/frame2.png \
  --duration <seconds> \
  --aspect_ratio 9:16 \
  --resolution 480p \
  --mode fast \
  --wait \
  --wait-timeout 30m

Run all segments in parallel using run_in_background: true (or your shell's equivalent).

If a segment 502s on first try, retry once. Higgsfield's queue is sometimes flaky.

Crossfade pattern (Stage 8)

./scripts/crossfade-segments.sh segA.mp4 segB.mp4 final_cold_cut.mp4 0.4
./scripts/extract-keyframes.sh final_cold_cut.mp4 7

Then read each keyframe with vision capability and produce a per-shot quality table for GATE 2:

Shot	Score	Notes
A-1	92	Character locked, push-in subtle
A-2	88	Slight eye flicker mid-shot
...	...	...

Hard limits to communicate upfront

Aspect	Cap (fast 480p)
Naturalness, AI-character	70–80
Mandarin lip-sync	70–85%
Multi-shot >5 cuts	65–80% precision
Character consistency cross-shot	80–90% (95+ with Soul ID)

Set realistic expectations BEFORE generation, not after.

Benchmark-Driven Adaptation Flow (high-leverage trigger)

When the user provides a benchmark URL (Instagram Reel / TikTok / YouTube Short) plus a topic in one message, run the full adaptation flow autonomously:

1. Run the analyzer: ./scripts/analyze-benchmark.sh <URL> ./benchmark 12 to download + extract 12 keyframes (4 hook zone, 4 body, 4 CTA zone)
2. Read the keyframes with vision capability; build a 9-lens benchmark report (hook mechanism, cut grammar, character consistency, energy curve, CTA placement, etc.)
3. Fetch the post page for caption + hashtag + CTA mechanism
4. Decide autonomously: visual style, N storyboard frames, M segments, hook mechanism, CTA strategy. Don't ask "do you want X or Y?"
5. Produce 3 artifacts: script.md, storyboard_prompts.md, prompt_segA.txt (and prompt_segB / prompt_segC if needed)
6. Run a marketing review via prompts/marketing-review.md (delegate to a marketing subagent if available, otherwise self-review)
7. Auto-adopt all proposed fixes — modify the artifacts directly, show the diff
8. Present GATE 1: hand back the benchmark analysis + adapted artifacts + adoption diff + estimated cost

Total wait time target: 3 minutes from URL to GATE 1.

For details, see docs/10-benchmark-adaptation.md.

When to use this skill

✅ Use when the user mentions:

• Seedance, Higgsfield, multi-shot video generation
• Character-consistent video series / IP work
• Vertical Reels (IG, TikTok, Shorts)
• Educational tutorials in video form
• Storyboard-based video pipelines

❌ Skip when:

• The user just wants a single static image (use a different image-gen skill)
• The user wants long-form video (>40 seconds — beyond Reel format)
• The user wants 16:9 horizontal cinema (different toolkit)
• The user is not using Higgsfield (the CLI specifics won't apply)

Anti-patterns to avoid

• ❌ Don't run higgsfield generate create nano_banana_2 ... — that's image generation, which is the human's job in the UI
• ❌ Don't ask "should we go cartoon or realistic?" — decide based on benchmark, present the call at GATE 1
• ❌ Don't skip GATE 1 — it's the cost frontier
• ❌ Don't skip Strict-Lock for educational content — character drift will ruin the take
• ❌ Don't generate at 1080p without explicit user request — find errors at 480p first
• ❌ Don't skip the marketing review at Stage 4 — even self-review catches conversion-killing issues
• ❌ Don't ask user "adopt all / some / reject?" after marketing review — default to full adoption, show the diff
• ❌ Don't write a script from scratch when the user has given you a benchmark URL — run the adaptation flow

Final note

The toolkit's whole reason for existing is to make IP video production repeatable. The user values consistency across episodes more than any single video being a masterpiece. Optimize for "this approach works again next week with a new topic," not "this one Reel is perfect."