screenplay-study-translation-skill

Screenplay Study Translation — Knowledge Reference

This file defines translation principles, extraction heuristics, and reference structure for the screenplay-study-translation project.

It does NOT define pipeline stages or AI behavior rules.

• •Pipeline stages → AI_AGENT_PROJECT_SPEC
• •AI behavior rules → AI_AGENT_CONTRACT

1. Translation Principles

• •Activation intent "生成预览" or "先给我第一批预览" means: initialize the

project, prepare required source/terminology/front-matter/reader-note artifacts, translate only the first batch, generate an HTML preview, and stop for user review.

• •Activation intent "翻译下一批" means: translate exactly one next batch and

stop after validation and preview unless the user separately authorizes continuous batch execution.

• •Activation intent "生成成品 HTML" means: merge validated batches, build the

final HTML, run audit, and stop when the reader HTML is ready.

• •Activation intent "导出 EPUB" means: export the final HTML as EPUB after the

final HTML audit passes.

• •Activation intent "预览通过，继续跑完整本" means: the first preview has been

accepted and the user explicitly authorizes continuous batch execution under AI_AGENT_CONTRACT.md; continue batch by batch until final HTML and EPUB delivery unless a required stop condition occurs.

• •Preserve the screenplay as a document, not only as dialogue.
• •When reference subtitles are provided, use subtitle content directly for

dialogue translations. Follow references/subtitle_alignment.md for label decisions.

• •AI translates non-dialogue elements: action, atmosphere, scene headings,

character cues, parentheticals, transitions, on-screen text, sound/effects, omitted scenes, and format markers.

• •When subtitles are absent, AI translates all screenplay elements.
• •Never replace action description with generic summaries.
• •Use the overall subtitle translation style as the primary input for

style-profile.json; apply that style profile to non-dialogue translation.

• •Subtitle quality is judged by the user. The skill parses and applies provided

subtitles, but does not automatically validate translation quality.

• •Require a user-supplied Chinese film title when initializing a new translation

project. Use it for reader-facing HTML title and cover text; do not infer it from filenames, subtitles, or model judgment.

• •Generate a terminology baseline automatically after source extraction and

subtitle normalization. Do not require routine user confirmation; escalate only high-impact ambiguity, conflicting evidence, or title-affecting terms.

• •For formal batch translation, prefer a current-range batch context package

when available; it is a token-saving read-only summary, not a validation gate or a replacement for batch validation.

• •When subtitles provide a proper noun translation, that translation takes

priority over AI judgment. It must be recorded in references/terminology.md and applied consistently across all batches.

• •When subtitles contain annotations or explanatory notes distinct from

dialogue translation, annotations must be included in the HTML output as reader notes, not dialogue. Never silently drop subtitle annotations.

• •If subtitles are absent, omit subtitle-specific labels.
• •When subtitles are absent, output quality depends on model judgment.
• •Preserve cinematic intent, dialogue rhythm, and cultural context.
• •Produce translation at final-quality intent from the first pass.

2. Extraction Principles

• •Reconstruct scene numbers from the source. Never invent sequential scene numbers.
• •Treat missing scene numbers as a script-type signal, not an automatic error. Reading drafts and public PDFs may omit production scene numbers.
• •Treat extraction rules as auditable heuristics, not universal truths. Do not turn one project’s formatting into a hard-coded global rule.
• •Keep page mapping explicit and consistent throughout extraction.
• •Auto-correct obvious, low-risk source defects when context makes the fix clear; record corrections in work artifacts.

3. Source Marker Handling

Independently scan and record:

• •Scene numbers
• •OMITTED scenes
• •CONT’D markers
• •MORE markers
• •Parentheticals
• •Transitions
• •Camera and format terms
• •Side-margin scene numbers

Stage 2 is a signal scan, not a rule-upgrade step. Record known markers as structural signals, unmatched structural evidence as warning signals, and low-confidence ignored candidates as noise signals. Do not create new marker types or schema fields from a single source scan.

Audit all markers against output before finalization.

4. Reader Assumptions

• •Assume readers may have limited source-language confidence.
• •Do not make user review the primary translation quality gate.
• •No raw screenplay terms should appear without translation or explanation.

5. Output Principles

• •Remove task labels, duplicate indexes, debug notes, and workflow artifacts from final reader output.
• •HTML must include a collapsible scene index with working navigation links.
• •Generated HTML should restore confirmed source-visible screenplay formatting

and structure after it has been captured in batch JSON, markers, layout metadata, or inline markup. Do not flatten source formatting signals into ordinary prose when source evidence shows they affect how the screenplay reads on the page.

• •Do not produce PDF output. PDF is deprecated in v0.3; EPUB is the supported

mobile reading output. scripts/export_pdf.py is retained only as historical reference.

• •Audit and final outputs are different presentations of the same formal translation: audit may show engineering metadata; final output hides technical noise.

6. Validation Checklist

This is the project-specific checklist referenced by AI_AGENT_CONTRACT Section 7.

Before finalizing, verify:

• •Stage 2 signal report produced before translation begins
• •source page mapping is explicit
• •source scene numbers are preserved as strings
• •OMITTED, CONT'D, MORE, split scene markers, and side-margin scene numbers are audited against output
• •subtitle labels are absent when no subtitles are provided
• •no raw screenplay terms appear without translation or explanation
• •HTML scene index links resolve
• •work batches live under work/batches/
• •final HTML and EPUB live under dist/

7. Must-Read References

Read only what the task needs:

Project flow:

• •references/workflow.md — end-to-end project workflow
• •references/configuration.md — project configuration
• •references/engineering.md — scripts and tooling

Source structure:

• •references/industry_conventions.md — unusual source structure, missing markers
• •references/marker_inventory.md — scanning and auditing source markers
• •references/source_lines.md — PDF text extraction and draft batches
• •references/validation.md — audit creation and execution

Translation and output:

• •references/batch_context.md — compact current-batch context packages
• •references/batch_schema.md — translation batch format and validation
• •references/terminology.md — screenplay, camera, and format terms
• •references/subtitle_alignment.md — subtitle alignment (when subtitles are present)
• •references/troubleshooting.md — unexpected behavior in extraction, translation, or HTML output