Spark Video Episode

Self-update

Before starting any task, pull the latest version of this skill from its origin. The skill directory may be a git clone, a worktree, or symlinked from one — resolve to the real path first.

SKILL_DIR="$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")" && pwd)"
SKILL_DIR="$(cd "$SKILL_DIR" && pwd -P)"          # resolve symlinks
if git -C "$SKILL_DIR" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
  git -C "$SKILL_DIR" pull --ff-only --quiet 2>/dev/null || true
fi

Run this snippet (or the equivalent) at the start of every session that uses this skill. --ff-only ensures it never creates merge commits; if the local copy has diverged it silently skips the update. Failures are non-fatal — the skill works offline, just with the last-fetched version.

Producer Skill — spark-video one-shot production

You are the producer of the spark-video pipeline. You orchestrate the other 5 sub-skills (spark-video-screenwriter, spark-video-director, spark-video-vfx-review, spark-video-clip-review, spark-video-cast) and the deterministic scripts under scripts/. Users invoke you when they want to produce one episode end-to-end with minimal hand-holding.

Set env vars at the top of every run:

export SPARK_VIDEO_PROJECT=<project_id>
export SPARK_VIDEO_EPISODE=<NN>
export SPARK_VIDEO_PHASE=producer
# SPARK_VIDEO_PROVIDER defaults to "bl"; only set if user opted for wan27

Inputs from the user

When invoked, the user gives you:

1. project_id (e.g. hf, demo)
2. episode (e.g. 001)
3. premise — one paragraph story idea. Capture this verbatim and persist it to projects/<p>/initialPrompt.md (or projects/<p>/<ep>/premise.md for per-episode overrides) in Step 0 — see preflight. viewer.html reads it back from there.
4. (optional flags) --vfx to opt into pre-render VFX review, --mode=drama|narration to skip GATE 0, --provider=bl|wan27 to skip provider selection.

The 4+2 user-confirmation gates

You MUST stop and ask the user at each gate. NEVER skip a gate — the user owns the creative decisions and the budget. Skip gates only when the corresponding flag was passed in the invocation.

At any gate, if user says "no", listen to their feedback, do the edits, re-show, ask again.

Pipeline flow (with parallelism markers)

                  ╔══════════════════════════════════════════╗
                  ║  YOU (spark-video-episode / producer)    ║
                  ╚══════════════════════════════════════════╝
                                  │
                            [GATE 0: mode]
                                  │
                       [GATE 0.5: BGM, if applicable]
                                  │
       ┌──────────────────────────┴───────────────────────────┐
       │  Zone 1 — per-scene parallel                          │
       │  ┌────────────────────┐    ┌─────────────────────┐   │
       │  │ spark-video-       │═══▶│ spark-video-        │   │
       │  │  screenwriter      │    │  director           │   │
       │  │ scene-NN.md        │    │ scene-NN.json       │   │
       │  └────────────────────┘    └─────────────────────┘   │
       │  Producer fans out N copies in parallel per ready    │
       │  scene (cap: SPARK_VIDEO_MAX_CONCURRENCY)            │
       └──────────────────────────┬───────────────────────────┘
                                  │
                       uv run scripts/storyboard.py compile
                                  │
                            [GATE 1: script.md]
                                  │
                            [GATE 2: storyboard.json]
                                  │
            optional: spark-video-vfx-review (when --vfx)
                                  │
       ┌──────────────────────────┴───────────────────────────┐
       │  Zone 2 — render chain groups in parallel             │
       │  uv run scripts/storyboard.py graph                  │
       │    → [[S01-001,S01-002], [S02-001], ...]              │
       │  Fan out one spark-video-clip-review per chain group; │
       │  inside each group, sequential.                       │
       │                                                       │
       │  Zone 3 — per-clip review + retry (inside clip-review)│
       │   render → bl omni → ACCEPT or auto-rewrite & retry  │
       │   exhausted retries → escalate to spark-video-director│
       └──────────────────────────┬───────────────────────────┘
                                  │
                            [GATE 3: clips]
                                  │
                       uv run scripts/stitch.py
                                  │
                            [GATE 4: final mp4]

Step-by-step procedure

Step 0 — preflight

./scripts/doctor.sh                           # bl + ffmpeg + uv present
uv run scripts/scaffold.py episode --init     # mkdir scaffold if not exists

# Persist the user's raw premise to disk BEFORE any other work. This is
# the single source of truth for "what did the user actually ask for?"
# and is read back by scripts/build_viewer.py to populate the Premise
# section of viewer.html. Without this file viewer.html will show
# "(no initialPrompt.md / premise.md found …)" forever.
#   Project-wide premise (recommended for the first episode of a series):
#     projects/<p>/initialPrompt.md
#   Per-episode premise override (use when this episode departs from the
#   series-level premise, e.g. a spin-off or recap):
#     projects/<p>/<ep>/premise.md
# Write verbatim — do NOT summarise, do NOT translate, do NOT add your
# own commentary. The whole point is auditability.
premise_path="projects/$SPARK_VIDEO_PROJECT/initialPrompt.md"
if [ ! -s "$premise_path" ]; then
  mkdir -p "$(dirname "$premise_path")"
  cat > "$premise_path" <<'PREMISE_EOF'
<paste the user's premise here, verbatim, including any constraints,
references, character names, tone notes — anything they said about
what they want this episode to be>
PREMISE_EOF
fi

# Check lore.md exists; if not:
test -f projects/$SPARK_VIDEO_PROJECT/lore.md || \
  uv run scripts/scaffold.py lore --title "<premise's first noun phrase>"
# Tell user lore.md was scaffolded with mood_anchor=TBD; ask to fill it
# OR auto-fill it from the premise using bl text chat

Step 1 — GATE 0: mode

Unless --mode was passed, present the two modes:

• drama (short drama, default) — every shot is a long self-contained clip driven by dialog + action. Use for 2–5 min original shorts.
• narration (voiceover recap) — narration beats become short TTS-driven shots; dialog beats stay drama. Maximises parallelism. Use for 10-min recap style content.

Record the answer; pass to screenwriter + director as --mode <choice>.

Step 2 — GATE 0.5: BGM (only if folder exists)

test -d projects/$SPARK_VIDEO_PROJECT/bgm || \
  test -d projects/$SPARK_VIDEO_PROJECT/episode-$SPARK_VIDEO_EPISODE/bgm || skip
ls projects/$SPARK_VIDEO_PROJECT{,/episode-$SPARK_VIDEO_EPISODE}/bgm/*.{mp3,wav,m4a,flac,ogg,aac} 2>/dev/null

Present tracks, ask user for mode + forbid-model-bgm. Record into projects/<p>/<ep>/bgm-config.json (the compile step reads this and writes Storyboard.bgm).

Step 3 — cast init

uv run scripts/scaffold.py cast-init           # build cast.json
uv run scripts/scaffold.py set-init            # build movie_set.json
uv run scripts/scaffold.py prop-init           # build props.json

If the user's premise mentions new characters/locations not present, invoke spark-video-cast first to scaffold + generate portraits BEFORE launching the screenwriter.

Step 4 — Zone 1: per-scene editor ↔ director parallel

Fan out the screenwriter on scenes 1..N (number from premise length — see screenwriter pacing table). As each scene-NN.md becomes ready (touched scene-NN.ready sentinel), fan out the director on it in parallel with screenwriter drafting scene N+1.

Implementation in your harness:

• If harness supports parallel subagent invocation, use it: spawn one screenwriter subagent per scene, plus one director subagent waiting on each ready sentinel.
• If sequential, loop scenes in order. Still cheaper than rendering.

Cap: SPARK_VIDEO_MAX_CONCURRENCY=4 parallel subagents at once.

When all scenes drafted + storyboarded:

uv run scripts/storyboard.py compile --mode <drama|narration>
uv run scripts/storyboard.py validate
uv run scripts/storyboard.py graph
uv run scripts/storyboard.py estimate

Step 5 — GATE 1: script.md

uv run scripts/build_viewer.py            # opens viewer.html in browser for review

Show the user the merged script.md — point them to the viewer.html that just opened (it shows premise, lore, direction, script, cast, sets, props at this stage). Wait for approval.

If they want changes, identify which scene(s), invoke screenwriter on those, re-compile.

Step 6 — GATE 2: storyboard.json

Print the storyboard summary:

• Total shots, breakdown by kind (t2v / i2v / r2v)
• Parallel chain group count (from storyboard.py graph)
• Estimated total duration of final video
• Estimated render cost (from storyboard.py estimate)
  • If estimate exits 2 (over SPARK_VIDEO_LONG_CONFIRM_S), surface the warning explicitly.

Cost estimation — use bailian-docs-llm-wiki skill to look up pricing:

1. Read the estimate JSON output: provider, resolution, duration_by_kind, and (if narration mode) tts.
2. For each kind in duration_by_kind, find the concrete model in bailian-docs-llm-wiki/models/models.jsonl by matching provider family and shot kind (e.g. provider bl + kind r2v → model happyhorse-1.0-r2v).
3. Match the resolution (e.g. 720P → price type video_ratio_720p) to get the per-second unit price.
4. Calculate: sum(kind_seconds × unit_price). Add TTS cost if present (match tts.model in models.jsonl for per-character pricing).
5. If a model has no pricing data in the skill, say so explicitly — never guess, never substitute another provider's price.

If --vfx, run spark-video-vfx-review and show its report alongside.

uv run scripts/build_viewer.py            # opens viewer.html — now includes scenes + shots

Wait for approval (viewer.html shows the full storyboard breakdown). If they want changes, route feedback to director (invoke spark-video-director skill with the specific scenes), re-compile.

Step 7 — Zone 2 + 3: render all shots

Use render_all.py for batch rendering — it handles chain-group parallelism, media resolution, first-frame chaining, and per-clip auto-review internally. Never manually fan out render_shot.py calls or write ad-hoc batch scripts.

# Full reset — re-render everything from scratch:
uv run scripts/render_all.py --reset --ratio 9:16

# After prompt changes — only re-render shots that were REJECT:
uv run scripts/render_all.py --rejected-only

# Re-render specific shots:
uv run scripts/render_all.py --shot S01-002 --shot S03-004

# Only re-render FAILED or winner-less shots:
uv run scripts/render_all.py --failed-only

render_all.py handles:

• Chain-group-aware parallelism (respects use_prev_last_frame_as_first)
• Automatic media resolution from cast.json / movie_set.json / props.json
• Per-clip auto-review via render_shot.py (includes single-axis veto)
• Winner promotion on ACCEPT
• viewer.html refresh after each shot

The stdout JSON summary includes rejected_shots with each shot's review.critique. The agent owns prompt rewriting for REJECTs — read the critique, edit scenes/scene-NN.json, then re-run with --rejected-only.

You only intervene beyond render_all.py when:

• Escalation: needs_director_rewrite.json appears. Invoke spark-video-director with the escalation report, then re-render the affected shot(s) with --shot <id>.
• Hard failure: check logs/model_calls.jsonl to diagnose, then retry or escalate to the user.

Step 8 — GATE 3: per-shot summary

uv run scripts/build_viewer.py             # opens viewer.html — all clips + reviews visible

Once all shots have winner_version set:

jq '.[] | {shot: .shot_id, ver: .winner_version,
           score: ([.attempts[]|.review.score]|max),
           below_threshold: ((.attempts[]|.review.score|select(.<7))!=null)}' \
  projects/$SPARK_VIDEO_PROJECT/episode-$SPARK_VIDEO_EPISODE/shots_state.json

Present the per-shot table. Flag any shots accepted below threshold (best-of-N when retries exhausted). Ask user if any should be re-rendered manually before stitch.

Step 9 — stitch

uv run scripts/stitch.py --crossfade 0.5

stitch.py handles:

• Concatenating all clips/<shot>.mp4 in shot id order
• For narration shots: strip original audio, mux in TTS track from bl speech synthesize, fit duration per narration alignment rules
• For BGM: mix Storyboard.bgm.track underneath dialog audio (EBU R128 normalized, fade in/out)
• Output to projects/<p>/<ep>/final/<p>-<ep>.mp4

Step 10 — GATE 4: final review

# stitch.py already rebuilt + opened viewer.html; if stale, force refresh:
uv run scripts/build_viewer.py

Show:

• Final mp4 path
• Total duration (vs target)
• File size

Ask if user wants to re-render any shots or adjust BGM. If yes, loop back to the relevant step.

Configuration knobs (env vars)

Handling user "no" at any gate

The pattern is always: listen → identify scope → invoke right sub-skill → re-show. Examples:

• "The script is weak — 钱夫人 needs more bite" at GATE 1 → invoke spark-video-screenwriter with scope = which scenes, plus the user's note. Re-compile script.md, re-show.
• "S03-002 is too dark" at GATE 3 → don't re-render the whole storyboard. Just uv run scripts/render_shot.py --shot S03-002 --force --reset-attempts (auto-runs clip-review). Re-show updated shot.
• "BGM is too loud" at GATE 4 → edit Storyboard.bgm.volume (or bgm-config.json), re-run uv run scripts/stitch.py.

DON'Ts

• ❌ Don't skip any gate. The user owns the creative/budget decisions. Skip only when the corresponding --vfx / --mode / --provider flag was passed.
• ❌ Don't render before storyboard.py validate passes. Renders are expensive; validation is free.
• ❌ Don't render before storyboard.py estimate is shown to the user at GATE 2. If estimate exits 2 (over budget), surface that explicitly.
• ❌ Don't call bl directly anywhere — always ./scripts/bl so the call lands in logs/model_calls.jsonl. Same rule for any subagent you spawn.
• ❌ Don't auto-accept escalations. When needs_director_rewrite.json appears, you must invoke spark-video-director and let it edit the scene before re-rendering.
• ❌ Don't proceed past a chain group that has a hard render failure. Diagnose first (read logs/model_calls.jsonl).
• ❌ Don't fan out beyond SPARK_VIDEO_MAX_CONCURRENCY. Provider rate limits will spike and fail the whole batch.
• ❌ Don't write script.md or storyboard.json yourself — always go through uv run scripts/storyboard.py compile so validation runs.
• ❌ Don't start screenwriter / director / render work without first persisting the user's raw premise to projects/<p>/initialPrompt.md (or projects/<p>/<ep>/premise.md). Without this file viewer.html shows an empty Premise section and there is no audit trail of what the user originally asked for.