菜单
Use when auditing Scinoephile Chinese conversion exclusions or checking whether Hans/Hant test subtitles change unexpectedly under OpenCC no-op direction conversions.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run and verify Scinoephile reference-guided transcription in consecutive block batches using the command-line interface, limited to audio extraction, transcription, delineation audit, and punctuation audit. Use when building or continuing a movie transcription from a provided media file and guide SRT without performing guided review or translation. Requires an explicit existing media file; never discover media from machine-specific locations.
Sync test fixtures in test/data/* by comparing the input and output files present with test fixtures in __init__.py and making appropriate updates, followed by updating the relevant tests to use the current fixtures/files. Use when test data file organization changes, output files are added/removed, or tests fail due to fixture mismatches.
Research and draft concise, operation-specific, prompt-ready additional context for LLM calls involving a movie, including canonical character names, aliases, places, objects, organizations, culture-specific terms, and other recurring terminology. Use when preparing or revising a Scinoephile movie dataset's additional_context for transcription, review, OCR, or translation; when an LLM needs title-specific naming or tone guidance; or when reference subtitles and web sources such as official film pages, Wikipedia, or Douban should be reconciled into a terminology list.
Audit character-aligned subtitle differences between a transcription and a non-authoritative reference, optionally using a timing-aligned guide for context. Use when evaluating transcription accuracy, explaining character-error-rate differences, distinguishing transcription errors from reference errors or acceptable variants, or inspecting aligned subtitle diff reports.
Audit Scinoephile OCR-fusion output against the exact two OCR source SRTs and optional OCR-fusion JSON used to generate it, optionally comparing every fused subtitle with a one-to-one validated SRT as ground truth. Use when inspecting ocr_fusion.json decisions, source selection, fused OCR errors, validation discrepancies, verification state, or corrected OCR-fusion cases.
Audit Scinoephile regular or guided subtitle reviews, traditional-to-simplified review paths, or dual Hans/Hant review paths. Use when inspecting original and reviewed SRTs, review or guided_review JSON, guide-supported revisions, character-focused changes, parallel script outputs, verification state, or corrected subtitle-review cases.
| name | audit-zho-conversion-exclusions |
| description | Use when auditing Scinoephile Chinese conversion exclusions or checking whether Hans/Hant test subtitles change unexpectedly under OpenCC no-op direction conversions. |
Use this workflow to verify that Chinese subtitle fixtures stay stable under no-op direction OpenCC conversions.
Run the bundled script from the repository root:
$env:UV_CACHE_DIR = "/tmp/uv-cache"
uv run python skills/audit-zho-conversion-exclusions/scripts/audit_zho_conversion_exclusions.py
Pass --root PATH when running outside the repository root.
The script discovers regular and OCR/PCR subtitle fixtures, audits Hans files with t2s, audits Hant files with s2t, filters active exclusions from scinoephile/lang/zho/script/conversion.py, and filters known inactive fixture artifacts encoded in the script.
Inspect each unexpected table row against the subtitle text and image dump link when one is available.
scinoephile/lang/zho/script/conversion.py only when the source text is legitimate and should remain unchanged in no-op conversion.Keep exception details in Python source rather than duplicating them in SKILL.md: active conversion exclusions belong in conversion.py, and known inactive audit artifacts belong in scripts/audit_zho_conversion_exclusions.py.
For the final response, paste the script output or briefly report the unexpected rows. For a clean audit, the script summary is enough.