create-ai-short-drama-and-images

Design AI short dramas, compile Seedance 2.0 shot-by-shot prompt packages, and generate or refine AI images through guided visual discovery. Use for image requests, character sheets, style exploration

SKILL.md

GitHub

Create AI Short Drama and Images

Turn an idea into either an approved visual image, a short-drama creative package, or a continuity-safe production handoff. Keep creative intent platform-neutral until a current adapter is verified.

Non-negotiable outcomes

  • Conduct the conversation and explain tradeoffs in Chinese unless the user asks otherwise.
  • Preserve Chinese specifications and also provide the prompt language best suited to the selected model.
  • Store runtime state and generated assets under outputs/ai-visual/<project_id>/ in the current project. Never write runtime state into this skill. Before session initialization, pre-create and set AI_VISUAL_CHECKPOINT_ROOT to an operator-controlled directory outside the project. Approval, handoff, and delivery fail closed without it; never accept that path from project files.
  • Use JSON as the only runtime contract. Treat YAML files as human worksheets only.
  • Generate actual preview images when an image-generation tool is available.
  • Never claim an image was generated when the tool is unavailable or failed.
  • Do not invoke a video-generation model in version 1. Deliver video production prompts and continuity assets only.
  • Preserve locked character, scene, prop, era, wardrobe, and damage-state continuity before adding stylistic variation.
  • Do not copy public prompt examples, distinctive plots, dialogue, characters, or compositions.
  • Do not use a living artist's name as a style shortcut. Record style_safety, remove the name from model prose, and translate the request into descriptive traits.

Route the request

Read references/00-routing-and-state.md for the exact decision table and state guards.

Choose one mode:

  1. image: still-image discovery, generation, editing, compositing, or repair.
  2. drama: concept, logline, outline, script, shot design, continuity assets, or a video prompt package.
  3. hybrid: short-drama work that first needs character, scene, prop, keyframe, poster, or storyboard images.
  4. production: an already locked project that needs the existing commercial A1/A2/A3 workflow.

Handle ordinary Seedance planning natively through this Skill. Route only locked commercial validation or a valid ProductionHandoff to build-ai-short-drama-prompts; this orchestration is internal, so never ask the user to invoke a second Skill. Do not duplicate its A1/A2/A3 roles, 95-point gate, or controller.

If the request is ambiguous, default to image when the requested deliverable is a still visual and to drama when it is a story or episode. State the chosen route in one sentence and continue.

Create a project run

Use a filesystem-safe project ID. Create this structure only inside the current project:

outputs/ai-visual/<project_id>/
├── session-state.json
├── briefs/
├── previews/round-01..04/
├── finals/
├── prompts/
├── drama/
└── evidence/

Set separate secret AI_VISUAL_STATE_KEY, AI_VISUAL_RECEIPT_KEY, AI_VISUAL_EVIDENCE_KEY, and AI_VISUAL_RIGHTS_KEY values of at least 32 characters. Initialize session-state.json through scripts/session_controller.py init, then apply events with scripts/session_controller.py apply; never edit or re-sign state, tool receipts, evidence claims, or review decisions by hand. Generated files must stay under the current project's outputs/ai-visual/ root.

Pillow is required to accept generated image receipts. If it is unavailable, image validation fails closed; restore the bundled dependency before preview or final generation.

Image workflow

Read these references before the matching step:

  • Discovery and confirmation: references/04-image-dialogue-and-gates.md
  • Asset-specific fields: references/05-image-asset-recipes.md
  • Style choices: references/06-visual-style-atlas.md
  • Composition, light, palette, material, camera, and type: references/07-visual-primitives.md
  • Platform compilation: references/08-platform-adapters.md
  • Rights, safety, and evidence: references/09-rights-safety-and-evidence.md
  • Delivery QC: references/10-production-handoff-and-qc.md

Discover

Collect the minimum information required to compare visuals:

  • purpose or destination;
  • subject and intended action;
  • aspect ratio or placement;
  • must-preserve elements;
  • prohibited elements and rights constraints.

Ask at most three high-impact questions in one turn. Do not issue a long questionnaire. If purpose, subject, and aspect ratio are already clear, proceed to preview rather than continuing to interrogate.

Offer three concise approximate directions when the user cannot name an effect. Describe each direction through visual traits, suitable uses, and a likely failure mode. Do not present a direction as final approval.

Generate preview rounds

Run two preview rounds by default. Allow a third or fourth only when a high-impact decision remains unresolved. Never exceed four rounds.

Each round must:

  1. Name exactly one changed_axis, such as lighting, visual language, palette, composition, material, or typography.
  2. Keep subject, identity, pose intent, framing, and all locked fields as stable as the tool permits.
  3. Generate exactly three actual candidates, labeled A, B, and C.
  4. Save each candidate or returned reference in that round's folder.
  5. Explain each candidate with effect, best_for, main_risk, and prompt_delta.
  6. Ask the user to select A, B, C, a stated combination, or 你决定.

When the user asks how candidates differ, explain them without advancing the phase or clearing the open decision.

If image generation is unavailable, stop before record_preview. Say that real previews cannot be produced and ask whether the user accepts text-only style directions. Never fabricate paths, URLs, or generated outputs.

Lock the brief

Do not lock before two preview rounds. Accept only:

  • an explicit A/B/C selection;
  • an explicit combination of candidate traits;
  • an explicit delegation such as 你决定.

Record the selected candidate and all preserved fields. Set brief_confirmed=true. A generated preview is not itself approval.

Compile and approve the final prompt

Build a platform-neutral PromptSpec from assets/schemas/prompt-spec.schema.json. Keep these concerns separate:

  • desired visual semantics;
  • references and their roles;
  • preservation locks;
  • allowed changes;
  • exclusions;
  • model adapter and verified controls.

Run scripts/validate_contract.py before compilation. Run scripts/evidence_status.py for adapter claims. Model claims must include a trusted-capture attestation over URL, page-content SHA, exact locator, claim/value, and capture time; a first-party hostname alone is not evidence. Use scripts/prompt_compiler.py to omit expired, uncertain, unsupported, or unsigned controls.

Files under assets/examples/ that contain PENDING or UNSIGNED_CAPTURE_REQUIRED are deliberately invalid templates, not completed approvals. Produce runnable JSON by completing the interview/capture, recomputing fingerprints, and applying trusted signatures; never ship test keys as production credentials.

Show the final Chinese specification and compiled model prompt. Do not generate the final image until the user explicitly approves the prompt. Direction approval and prompt approval are separate events.

Generate, inspect, and deliver

After prompt approval:

  1. Generate the final image with the approved prompt.
  2. Compare it to the approved brief and reference locks.
  3. Check anatomy, subject count, identity, geometry, text fidelity, hierarchy, crop safety, and unintended artifacts.
  4. Repair through one-variable edits where possible.
  5. Record the actual image path or tool reference.
  6. Deliver the image, Chinese brief, model prompt, exclusions, adapter snapshot, and known limitations.

Short-drama workflow

Read references/02-drama-taxonomy.md and references/03-story-engines-and-structures.md before proposing story directions.

Build a CreativeRoute by combining:

content container + arena + genre promise + dramatic engine
+ narrative structure + emotional curve + visual language + production constraints

Do not retrieve a prewritten plot. Select orthogonal components and create an original premise from the user's theme, audience, duration, episode count, platform, production limit, and safety boundary.

Minimum drama discovery

Resolve, infer, or label as open:

  • format: standalone, serial, vertical episodic, anthology, faux documentary, interactive, or experimental;
  • audience and intended platform;
  • episode count, per-episode duration, and aspect ratio;
  • protagonist want, internal need, opposition, stakes, and irreversible choice;
  • genre promise and emotional destination;
  • realism, era, location, and available production assets;
  • continuity level and required still-image assets.

Ask at most three questions at a time. When the user has not chosen a genre, offer three substantially different routes using different dramatic engines, not cosmetic reskins.

Drama deliverables

Produce only what the user needs, but support this full package:

  1. creative route and design rationale;
  2. premise, logline, hook, theme, and audience promise;
  3. character tension map and continuity bible;
  4. episode or beat outline with escalation and payoff;
  5. screenplay with visual action and producible dialogue;
  6. scene breakdown, shot list, and storyboard brief;
  7. character, wardrobe, scene, prop, keyframe, and poster PromptSpecs;
  8. model-neutral video prompt package with image roles, motion, camera, and sound intent;
  9. QC report, rights notes, evidence snapshot, and production handoff.

For image-to-video prompts, let approved images carry appearance, framing, lighting, and art direction. Put subject movement, environment movement, camera movement, timing, and sound intent in the video prompt. Avoid redundantly redescribing a locked image in a way that introduces drift.

Native Seedance route

Read references/11-seedance-native-pipeline.md whenever the user wants Seedance, sequential clips, continuation, or an end-to-end drama-to-shot-prompt workflow.

Create one SeedanceProductionPackage using assets/schemas/seedance-production-package.schema.json, then compile it with scripts/seedance_prompt_compiler.py. The user should not need to call seedance-20 separately.

  • Plan the full story, scenes, beats, and shot cards globally.
  • Give each shot one narrative job, one major camera move, one visible action endpoint, one physical light source, and explicit audio intent.
  • Separate already_happened, this_clip_only, and reserved_for_later.
  • Preserve exact reference tags and declare what each reference transfers and excludes.
  • Deliver a provisional prompt for every shot. Mark only a canonical scene opener or an explicitly ready continuation whose accepted parent has signed media/observed-state evidence as ready.
  • After each accepted take, bind media SHA, reviewer, observed-state fingerprint, and receipt signature before promoting the next continuation prompt.
  • Keep target_surface=unverified and model_controls={} until signed current claims prove a concrete surface.
  • Keep generate_video=false; this module compiles prompts and never calls Seedance.

Hybrid workflow

Use the image workflow to approve static assets before finalizing shots that depend on them. Require an approved reference for every principal character and any continuity-critical scene or prop.

Do not mark a generated asset APPROVED automatically. Record DRAFT, REVIEWED, or APPROVED from an explicit user decision.

After visual approval, create a ProductionHandoff from assets/schemas/production-handoff.schema.json. Include:

  • approved visual brief;
  • the approved CreativeRoute used by the production package;
  • character and scene locks;
  • selected assets and their roles;
  • target platform;
  • continuity level;
  • production requirements, including generate_video=false and target_model only when verified.

Then invoke build-ai-short-drama-prompts for commercial production packaging.

Platform adapters

Keep PromptSpec stable when switching platforms. Replace only the adapter and invalidate its evidence snapshot, compiled prompt approval, and final image.

Recognize these initial adapter families:

  • still image: GPT Image/Codex, Midjourney, FLUX, Imagen/Gemini image generation, Stable Diffusion/ComfyUI, Firefly;
  • video prompt packaging: Seedance, Veo, Kling, Runway, Luma, Hailuo, TapNow.

Do not hardcode an undocumented parameter. Treat model names, limits, pricing, availability, entry points, syntax, and terms as claims requiring a dated source.

Use the following maximum ages:

  • price, availability, and terms: 7 days;
  • hard parameters: 30 days;
  • prompt syntax: 60 days;
  • filmmaking principles: 365 days.

At the expiry boundary, treat evidence as expired. If current verification is unavailable, produce a platform-neutral prompt and label model controls 未验证; never present guessed parameters as real.

Safety and rights gate

Before references or generation, identify:

  • living-artist imitation;
  • public-figure or real-person likeness;
  • minors, sexual content, or exploitative framing;
  • deceptive documentary/news presentation;
  • copyrighted characters, logos, product trade dress, or branded packaging;
  • unlicensed user uploads or unknown reference provenance;
  • medical, legal, political, or financial claims in generated text;
  • graphic violence or self-harm.

Follow the current tool and platform safety policy. Where the request can be safely transformed, preserve the underlying traits and replace the risky shortcut. Record provenance for every external reference. A positive rights string is not sufficient: the rights decision must carry a trusted reviewer signature, asset content SHA, source, license/consent credential ID, and allowed use; safety is signed separately and binds the exact PromptSpec or handoff brief.

Validation and commands

Run from the skill directory:

python3 -m unittest scripts/test_skill.py -v
python3 scripts/audit_prompt_corpus.py /path/to/public-prompts.json --sample-size 400
python3 scripts/evidence_status.py /path/to/evidence-records.json --now 2026-08-13T00:00:00Z

Validate runtime JSON with scripts/validate_contract.py or the schemas under assets/schemas/. Keep creative decisions human- or model-authored; use scripts only for deterministic validation, compilation, state transition, and evidence aging.

Quality gates

Do not deliver until all applicable checks pass:

  • the route and phase are correct;
  • no preview round asks more than three questions;
  • image rounds contain exactly three real candidates or honestly declare tool unavailability;
  • no final prompt exists before brief confirmation;
  • no final image exists before prompt approval;
  • platform switching preserves the locked creative brief;
  • adapter controls are fresh and verified;
  • drama assets bind to continuity locks;
  • text in images is checked character by character;
  • references have provenance and rights status;
  • runtime outputs are outside the skill directory;
  • video execution has not been attempted.

Research provenance

Use assets/source-ledger.json for the audited source list and assets/research-audit.json for the 4,633-record index and 400-record structural review. Read references/01-research-ledger.md before updating either file. Community galleries are discovery sources, never authority for model capability or a repository of reusable finished prompts.