Phantomlau3674/ai-film-director-workbench

AI Film Production System

This skill owns project truth and production gates. It does not write final AI video prompts in chat. Final prompt text comes from the isolated ai-video-prompt-generator LLM generator.

Required Reading

• references/v2-production-protocol.md before changing project state.
• references/manifest-contract.md before editing manifests.
• references/theory/README.md before changing film-theory production diagnostics, review checks, or theory-linked failure taxonomy.
• references/ai-drama-production-hypotheses.md before designing, optimizing, or gating Douyin-style AI dramas, AI microdramas, AI manju/manhua-video, serial AI shorts, or public-case methodology migrations.
• references/director-dense-prompt-production-method.md before approving complete timed prompts with micro-actions, camera geometry, visibility, continuity, transition, dialogue, sound, static action-tableau, cross-video handoff, representation-layer, or semantic-color contracts.
• references/storyboard-blender-seedance-workflow.md before designing or reviewing storyboard-board, Blender/blockout previs, first/last-frame, control-pass, camera-path, or Seedance/Dreamina/Jimeng-style workflows.
• references/director-grade-asset-workflow.md before planning a multi-shot film's character plates, environment plates, scene director packs, shot control packs, or multimodal provider selections.
• references/asset-prompt-authoring-method.md before writing or reviewing prompts for character, costume, environment, prop, creature, vehicle, palette, storyboard, staging-map, start/end-frame, or static action-key assets.
• templates/manifest.schema.json, templates/assets.schema.json, templates/constraints.schema.json, templates/shot_plan.schema.json, and bible/provider schemas when changing validators.

Project Contract

One movie root contains:

_project/
  PROJECT.yaml
  CURRENT.yaml
  manifests/
  decisions/
  audits/
  constraints/
    film_theory_contract.default.json
    scorecard_theory_rules.default.json
  episode-plans/
  task-board/
  failure-taxonomy/
  providers/
01_script/
  shot_plans/
02_assets/
  refs/
  registry/
    assets.jsonl
    audio_assets.jsonl
  bibles/
    characters/
    scenes/
    props/
  audio/
  subtitles/
03_prompts/
  specs/
  generated/
  preflight/
  lint/
04_jobs/
  dreamina/runs/
  providers/runs/
05_review/
  trackers/
  diagnostics/
  scorecards/
06_edit/
  playlists/
  timelines/
  releases/
  exports/
99_archive/

v2 Workflow

1. Apply the asset necessity gate, then register only justified assets with exact duties. Every registered asset needs a named production job and at least one consumer (used_by) or reuse scope. Do not scaffold every asset family for an exploratory or quiet shot:

python scripts/register_asset.py C:/movie/film-root \
  --path 02_assets/refs/scenes/example.png \
  --asset-type scene \
  --asset-kind environment_master \
  --priority primary \
  --visual-duty environment_geography \
  --role "Recurring location master" \
  --production-job "Lock connected-room geography, scale, materials, and motivated light" \
  --must-lock "entrances, exits, walkable paths, material zones" \
  --must-not-lock "shot-specific cast blocking, storyboard arrows, final camera path" \
  --reuse-scope "SCENE_04 all shots"

2. Parse script text into an episode-plan scaffold when the task starts from a script file:

python scripts/build_episode_plan.py C:/movie/film-root \
  --script 01_script/script.md \
  --output _project/episode-plans/episode_plan.json \
  --set-current

3. Capture constraints, task queue, and bibles before shot specs when the task starts from script or outline.

   For theory-informed projects, record the film-theory contract as production intent, not as art-school compliance. It may define reality mode, image regime, genre promise, spectator-position audit policy, and waivable diagnostic gates.

   For case-informed AI drama projects, record the scene or episode promise, first visible tension, continuity engine, anchor duties, final hold/handoff, differentiation note, and platform-risk notes as project truth before prompt generation.

   For storyboard/Blender/Seedance workflows, register storyboard frames, blockout videos, control passes, camera metadata, and reference media duties before prompt generation. Treat Blender/previs as a control source, not as story truth.

   Also name the complexity owner: prompt_owned, asset_delegated, existing_video_delta, or dialogue_audio. Long text is appropriate when text owns multi-beat execution. Compact text is appropriate only when an approved asset or source clip owns the omitted detail.

   When assets must be created, author each asset from its downstream production question. Record must_lock, must_not_lock, reuse scope, invariants, and consumer shots. Character sheets, environment masters, staging maps, prop sheets, palettes, storyboards, and clean start/end frames are different control sources; do not ask one decorative image to replace all of them.

4. Build a shot-plan scaffold when the user has not already split shots:

python scripts/build_shot_plan.py C:/movie/film-root \
  --episode-plan _project/episode-plans/episode_plan.json \
  --output 01_script/shot_plans/shot_plan.json \
  --set-current

5. Track async work in the task queue:

python scripts/manage_tasks.py add C:/movie/film-root \
  --layer prompt --title "Generate prompt for SHOT001"
python scripts/manage_tasks.py next C:/movie/film-root --layer prompt

6. Validate project layers and prompt specs:

python scripts/validate_project_layers.py C:/movie/film-root
python scripts/validate_preflight.py C:/movie/film-root \
  --spec 03_prompts/specs/SHOT.prompt_spec.json

7. Generate prompt output with ai-video-prompt-generator.

8. Build or update a manifest row pointing to:

• shot_spec_path
• prompt_output_path
• registered reference_images
• provider, generation, audio/subtitle, and review state

9. Prepare a provider-neutral or Dreamina batch only after preflight passes:

python scripts/prepare_provider_batch.py C:/movie/film-root \
  --manifest _project/manifests/manifest_v2.jsonl \
  --batch-id YYYYMMDD-HHMMSS_batch

python scripts/prepare_dreamina_batch.py C:/movie/film-root \
  --manifest _project/manifests/manifest_v2.jsonl \
  --batch-id YYYYMMDD-HHMMSS_batch \
  --slots SHOT001

10. Build timeline/export packages only from reviewed clips:

python scripts/build_timeline.py C:/movie/film-root \
  --manifest _project/manifests/manifest_v2.jsonl \
  --timeline-id rough_cut
python scripts/prepare_ffmpeg_export.py C:/movie/film-root \
  --timeline 06_edit/timelines/rough_cut.json \
  --export-id rough_cut

Gates

Set the project or prompt validation stage explicitly. During exploration and proof_shot, the director contracts below are modular guidance and unresolved items are warnings unless they make the request unsafe or structurally invalid. During batch and release, relevant unresolved items become blockers. Do not require a contract that has no bearing on the shot's actual ambiguity or risk.

• No prompt spec: do not generate.
• Invalid prompt spec: do not generate.
• Missing reference asset registration: do not batch.
• Missing or malformed constraints, episode plan, provider profile, task board, or bibles: fix the project layer before production.
• High-motion batch/release spec without action_arc: do not generate.
• Crowd/battle/sports/dance/mount spec without background_motion_layers: do not generate.
• High-motion spec without start/end/action reference or waiver: do not generate.
• Missing editability contract: warn during exploration/proof; block batch and release.
• Case-informed AI drama without scene/episode promise, first visible tension, continuity engine for multi-shot work, anchor duties, or platform-risk note: do not batch beyond isolated proof shots.
• Public-case methodology without source date, provenance level, and skin removal: do not promote it to reusable skill guidance.
• Public-case methodology without failed-sample comparison: do not describe it as a formula, rule, or causal explanation.
• Director-dense prompt without contiguous time coverage and complete micro-action chains: do not batch.
• Compact multi-beat prompt without an approved controlling storyboard, start/end-frame pack, reference pack, source video, or explicit complexity owner: do not batch.
• Director-dense prompt with continuous-take/cut, allowed/forbidden camera, visibility/reveal, music/silence, or reference-usage conflicts: surface the conflict during exploration; resolve or waive it before batch generation.
• Do not reject a director-dense prompt merely because it is long. Reject it when clauses have no control duty, required local actions are missing, or the prompt is physically contradictory.
• Do not treat a provider SUCCESS status as creative acceptance. Record review, instruction adherence, artifacts, editability, and repair history.
• Do not turn lexical keyword collisions or unverified metadata-field differences into hard failures without scope and product-semantics review.
• Provider submission over a verified model/version capacity limit: compile a shorter provider prompt from the saved director contract before submission; preserve state chains, camera landing, reference duties, and handoff.
• Action-tableau storyboard with ambiguous visible/hidden anatomy, unsupported containment/contact, or no visible causal/environmental evidence: do not use as a start frame.
• Multi-shot clip that changes physical/inverted/stylized/schematic/media layer without a representation-layer contract, or continues a prior clip without an explicit handoff contract: do not batch.
• New prompt or Blender/previs workflow without saved run evidence: mark it experimental and do not use it as a batch gate.
• Complex action, multi-person blocking, or camera-path-dependent shots without visual_staging_contract, registered control media, or an explicit waiver: do not batch.
• Blender/previs requested only for "cinematic feeling" without a control duty: do not add Blender work.
• Prompt output not linted: do not submit to Dreamina.
• Cancelled task: do not resume without an explicit manage_tasks.py resume.
• Review not accepted: do not enter final release.

Film Theory Layer

Film theory is used as a production diagnostic layer:

• episode plans may record narrative design, scene turn, knowledge delta, and causal links;
• shot plans may record viewer question/answer, camera motivation, visual axis, blocking, edit contract, sound contract, and AI risk mitigation;
• scorecards may include observable theory-linked checks;
• failure taxonomy may link symptoms to theory cards and repair actions.
• AI drama diagnostics may record visible tension, promise payoff, continuity-engine health, anchor-duty completeness, final hold/handoff legibility, differentiation from trend skin, and platform risk.

Do not turn theory into a universal hard gate. Hard stops belong to production truth: invalid schemas, missing assets, missing bibles under strict mode, broken identity, broken action causality, missing editability, unsafe or unlicensed release context, and rejected review. Theory guidance should usually be warning-level, review-level, or waived with an explicit project reason.

Responsibility Split

• This skill owns project files, assets, manifests, gates, batches, review status, scorecards, timelines, playlists, releases, and film-theory production diagnostics.
• ai-video-prompt-generator owns prompt spec validation, isolated LLM prompt generation, and prompt-output linting.
• dreamina-cli owns external generation, query, download, credits, and run ledgers.