AI Short Drama Director

AI Short Drama Director

Overview

Run a staged AI short-drama production pipeline. Preserve the story while turning it into stable, reviewable artifacts for GPT Image 2 and Seedance generation.

Start Every Project

1. Read references/00-workflow.md.
2. Inspect the project folder before generating anything. Reuse existing scripts, character images, scene images, prop images, effect images, plans, and Segment packages.
3. If the project has not been initialized, run:

python3 scripts/init_project.py <project-folder> --episode 第一集 --mode confirm-each

4. Ask the user to choose one execution mode:
   • confirm-each: pause after every stage and offer both 确认后继续 and 从这里自动完成.
   • auto-from-current: lock accepted artifacts and finish all remaining stages automatically.
5. Even in automatic mode, pause on any hard-stop condition defined in references/00-workflow.md.

Run the Pipeline

Execute stages in order. Resume from the first incomplete stage when artifacts already exist.

Apply Global Rules

• Preserve plot direction, character relationships, scene settings, key props, and every original spoken-text string.
• Treat dialogue, inner monologue, voice-over, narration, and authored subtitles as immutable source text. Preserve their wording, pronouns, sentence order, punctuation, repetitions, hesitations, and code-switching verbatim unless the user explicitly authorizes a rewrite.
• Add visible performance detail, spatial clarity, and camera logic. Do not add major events, reversals, relationships, or character-changing behavior.
• When generating Seedance prompts or prompt columns, enforce the director reasoning gate in references/06-seedance-delivery.md. Do not create final prompts by template filling, row-field paraphrase, keyword expansion, or unattended batch scripting. Each final prompt must be reasoned from the source script beat, a fixed scene floor plan, shot-appropriate camera design, visible screen composition, character intention, and dramatic purpose.
• Limit each Segment to at most 15 seconds. Permit a shorter final Segment at the end of a scene. Start every new scene in a new Segment.
• Keep shot-to-shot and Segment-to-Segment continuity explicit. End Segments on stable framing or completed action whenever possible.
• Use concrete names instead of ambiguous Chinese pronouns in action descriptions, blocking, asset mappings, and other production instructions. Never apply this rule inside dialogue, inner monologue, voice-over, narration, or authored subtitles.
• Prefer existing visual assets. Generate only missing identities, missing views, or missing execution-critical details.
• Use GPT Image 2 for generated visual assets. If GPT Image 2 is unavailable, pause and tell the user; do not silently substitute another model.
• When creating spreadsheet outputs, use the available spreadsheet capability and deliver both .xlsx and HTML.
• Auto-select the strongest visual candidate in automatic mode and keep rejected candidates in 淘汰图/ for review.

Validate and Package

Before delivery:

1. Before accepting Stage 02, run the dialogue fidelity check against the original source and the director-ready script:

python3 scripts/check_dialogue_fidelity.py <original-script> <director-ready-script>

Stop and restore the original wording if this check reports any missing or changed spoken text. Do not continue downstream with a failed Stage 02.

2. Verify that the XLSX and HTML tables represent the same rows and required columns.
3. Run prompt checks:

python3 scripts/check_segment_prompts.py <project-folder>/06-Seedance提示词

4. Create each Segment package from a JSON manifest:

python3 scripts/package_segment.py <manifest.json> <project-folder>/07-Segment素材包

5. Check that asset order matches @Image1, @Image2, and subsequent references in that Segment prompt.
6. Report reused assets, generated assets, automatically selected images, rejected images, prompt-length warnings, and any unresolved hard stops.

Use Resources Selectively

• Read only the reference file needed for the current stage.
• Use scripts/init_project.py to create a predictable project structure and workflow state.
• Use scripts/check_dialogue_fidelity.py to block any unauthorized spoken-text rewrite at Stage 02.
• Use scripts/check_segment_prompts.py to enforce prompt independence and delivery constraints.
• Use scripts/package_segment.py to copy and rename Segment assets deterministically.