Ai video prompt generator

AI Video Prompt Generator

This skill no longer writes final prompts directly inside the Codex chat context. It consumes a clean prompt_spec.json, validates it, calls an isolated LLM prompt generator, and lints the output.

Required Reading

Read only what the task needs:

• references/v2-core-protocol.md for the generic v2 prompt contract.
• references/prompt_spec.schema.json before creating or changing specs.
• references/output-contract.md before consuming generator output.
• references/seedance-provider-prompting.md when the target provider is Dreamina, Seedance, Jimeng, or another Seedance-like multimodal video surface.
• references/director-dense-seedance-protocol.md when the requested clip is a continuous take, timed multi-beat sequence, dialogue scene, transformation, occlusion transition, action tableau, cross-video continuation, representation-layer change, or other director-dense prompt.
• references/asset-duty-handoff.md when the prompt uses character, environment, prop, palette, storyboard, staging, frame, source-video, voice, or cadence assets.
• director_profiles/*.md only for the selected director profile.

Do not load old project examples, deprecated compatibility rules, or chat history into the generator. Project-specific examples belong in the movie root, not inside this generic skill.

Workflow

1. Build or receive a prompt_spec.json from project truth. Choose validation_policy.stage before adding detail: exploration for creative discovery, proof_shot for a runnable experiment, batch for repeatable production, or release for final delivery. Missing creative contracts are warnings in the first two stages and blockers in the last two. Legacy specs without this field remain batch for compatibility. Choose prompt_density_policy.mode deliberately: compact_motion for one controllable event, or director_dense for a complete timed execution contract. Long output is valid in director-dense mode; completeness, not brevity, is the gate. Also choose prompt_density_policy.control_source: prompt_owned, asset_delegated, existing_video_delta, or dialogue_audio. A compact prompt may carry several inherited beats only when approved assets or a source clip own those beats and the spec names the inherited and text-owned controls. Do not expand a one-delta edit or voice cue into a scene rewrite. Treat director contracts as modules, not a form to fill. Use only the modules that reduce a real ambiguity or known failure risk for this shot. Consume approved asset duties; do not paste upstream character/scene/prop generation briefs into the video motion prompt. Canonical asset creation remains with ai-film-production-system; this skill binds those assets to the shot and protects them from cross-duty contamination. If existing_video_delta or dialogue_audio already has an authoritative source picture, skip still/storyboard generation. An empty storyboard_prompt is then correct control allocation, not missing work.
2. Run:

python scripts/validate_prompt_spec.py path/to/prompt_spec.json

3. Only when the shot needs a new still storyboard, keyframe, reference image, or start frame, generate the image prompt with the dedicated storyboard generator. Skip this step for source-owned picture work:

python scripts/generate_storyboard_prompt.py \
  --spec path/to/prompt_spec.json \
  --director high_action_physics \
  --output-dir path/to/output

Use the resulting <shot_id>.storyboard_prompt.txt for image generation. This stage must not produce motion_prompt or prompt_text. Before the LLM call, the generator projects the production spec into still-image truth: a neutral start frame receives the frozen starting composition, POV geometry, visible subjects/props, positive style, and reference duties, while action arcs, background motion, end states, sound, failure lists, and negative policy remain outside the still prompt. action_tableau keeps only its static topology and causal evidence.

4. For image-to-video, multi-image video, or text-to-video motion, generate the video motion prompt with the dedicated motion generator:

python scripts/generate_motion_prompt.py \
  --spec path/to/prompt_spec.json \
  --director high_action_physics \
  --storyboard-prompt path/to/output/SHOT.storyboard_prompt.json \
  --output-dir path/to/output

Use the resulting <shot_id>.motion_prompt.txt for video generation. This stage must not rewrite the still image prompt. Omit --storyboard-prompt when the source video/current frame already owns the picture.

5. For adapters that still need one combined compatibility payload, generate prompt output:

python scripts/generate_prompt.py \
  --spec path/to/prompt_spec.json \
  --director high_action_physics \
  --output-dir path/to/output

For Seedance/Dreamina-like providers, generate_prompt.py internally uses separate storyboard and motion LLM passes. Use --dry-run for deterministic local checks without API calls.

6. Lint the combined prompt output:

python scripts/lint_prompt_output.py \
  --spec path/to/prompt_spec.json \
  --prompt-json path/to/output/SHOT.prompt.json

Environment

generate_prompt.py uses the OpenAI SDK:

• OPENAI_API_KEY is required unless --dry-run is used.
• PROMPT_LLM_MODEL is required unless --model or --dry-run is used.
• PROMPT_LLM_TEMPERATURE defaults to 0.6.
• For OpenAI-compatible providers, pass --base-url and --api-key-env, or pass --env-file and use a provider-prefixed model such as volcengine/doubao-seed-2.0-lite.
• Volcengine Coding Plan defaults to https://ark.cn-beijing.volces.com/api/coding/v3 when the model prefix is volcengine/, doubao/, or ark/; it reads VOLCENGINE_API_KEY.
• MODEL_NAME is accepted as a fallback when PROMPT_LLM_MODEL is not set.

Contract

The generator may use only the spec and selected director profile. It must not:

• invent story beats not present in the spec;
• read movie folders, manifests, skills, or chat history;
• merge multiple smallest events into one prompt;
• turn embodied POV into absent POV;
• omit action arcs, background motion layers, or reference duties that the spec supplies.

For director_dense, preserve every supplied temporal beat and micro-action detail. In exploration/proof-shot work, do not invent empty contracts or force all possible sections. In batch/release work, unresolved required contracts are blockers. A long prompt is not a failure by itself.

For asset_delegated, preserve the storyboard/start-frame/reference pack as the controlling source and write only the unresolved motion, continuity, or sound controls. For existing_video_delta, name the source video, requested delta, and preservation rule; do not restage unrequested details. For dialogue_audio, bind line, speaker, voice, delivery, timing, and visible reaction without inventing unrelated blocking.

For storyboard_frame_mode: action_tableau, preserve body orientation, visible/hidden anatomy, contact/containment, prop state, pose dynamics, causal traces, environmental motion evidence, narrative read, and interpretation guards. Do not flatten a partial-body pose into a generic standing/sitting pose.

Codex remains the coordinator: it prepares specs, validates them, calls the generator, lints the result, and hands approved prompt outputs to production. Prompt JSON must keep prompt_text, negative_prompt, audio_prompt, risk_notes, safety_notes, and reference_duties as separate fields.

For Dreamina/Seedance-style providers, prompt output is split:

• storyboard_prompt generates still storyboard/reference/start-frame images.
• motion_prompt moves that image in video generation.
• prompt_text remains a video-provider compatibility field and should equal motion_prompt.

Do not paste internal rationale, failure taxonomies, or unbounded negative prompts into provider-facing fields. Director-facing execution sections are allowed when they directly control or verify the generated clip.