VideoLingo prompt provenance

Upstream: https://github.com/Huanshere/VideoLingo
Revision: 01f843492759be9bbfd1a5cb697b799cc9de1609
File: core/prompts.py
SHA-256: 581370e1c91f90285015bc50709ead89f3a51d95e9461fd40ac4cddf9d1c048b
License: Apache-2.0 (full text in ../assets/VideoLingo-LICENSE).

The original nine template families adapt all nine prompt-producing functions in that source file. Python f-string substitutions become named input slots. The original roles, tasks, translation principles, review criteria, split alternatives and trim examples are retained. align.txt now extends the single alignment into two candidates as described below. Expressiveness requests brief editorial findings instead of private reasoning. Bindings and validation notes describe VidBee-specific tool, storage and timestamp constraints.

Orchestration also follows core/_4_1_summarize.py, core/_4_2_translate.py, core/translate_lines.py and core/_5_split_sub.py at this revision: theme plus glossary; roughly 600 source characters or 10 lines per batch; previous 3/next 2 lines of context; direct then optional reflective translation; at most 3 attempts; faithful-result fallback for failed polishing; matching nonempty split/alignment parts.

subtitle-constraints.txt adapts the actual segmentation, display-length and timing policies beyond the prompt templates. The inspected local checkout is at the same revision; these constraint sources have no working-tree changes. SHA-256:
  config.yaml: 2967eee3d9f7a59fb7b42ec77905fb2068e2e098b500707c92e3bd6dd3091273
  core/_3_2_split_meaning.py: 361cbbe8aebdc723f57ee84121d1ef1123a1f2e547b7b7f06c58119d6a0727a0
  core/_4_2_translate.py: cb9f410ed4553976c1385786f907953843fc09abfcb744b2decc599b49895be5
  core/_5_split_sub.py: dde160b6ee9fbed445bdd5adb5a194c715e2458488a4793a615a4464e454d8cd
  core/_6_gen_sub.py: 222402f1401a5ee50ad81fe1cc726ea8d5ffed404afe053d8b306896da082c3d
  core/_8_1_audio_task.py: abd6e2523e9a890ffff3a7937e1c211f8ddbebf28500f88d9c4729d9f397cec1
  core/_8_2_dub_chunks.py: 581b35a47cb290a6689aa60779a7901d56f8db92821dd68e984ee0d1f591964f
  core/spacy_utils/split_by_mark.py: 6bac388cb19e1d4cdef0f4da7b1b5456e5c012f5d4cd5dbde230022c2e104e1f
  core/spacy_utils/split_by_comma.py: 2a0a2ce5a30d0a27437e799314822d560f0f6bca0bc3ca1f169134c32141f412
  core/spacy_utils/split_by_connector.py: 51da295940eedf343b4c5aec19bef7b8e6ded601f429747bd52305dde89e1de6
  core/spacy_utils/split_long_by_root.py: 2dccafa27a64ebc6569a4aabd775ee1e0d890e7860e3d858c535a367e3762a16

The policy preserves the 20-word semantic target, 75-character source budget, 1.2 target multiplier, implemented Unicode weights, two-way split alternatives, mechanical fallback, matched bilingual parts and at most three split/recheck rounds. It distinguishes these from translation batching and prompt-validation retries. The stale calc_len comment is not used as the weight definition.

VidBee adaptations are explicit: split parts normally become display rows inside unchanged cue intervals; new intervals need exact source word timing; saving the target does not reflow the read-only original; unresolved widths are handed back instead of being treated as validated. Upstream sub-second gap filling, 2.5-second audio cue merging/extension and 1.5-second chunk tolerance are documented but not applied. The upstream 3.5-second conditional voiceover trim and 1/1.2/1.4 speed settings are audio policy, not display requirements; VidBee keeps its actual dubbing-tool fit checks and supported correction paths. These instructions do not add programmatic subtitle-width enforcement.

VidBee uses native skill reads, existing notes, transcript timestamps and add_subtitles. It does not run the upstream Python application, Excel intermediates or its task runner. All upstream brand-role wording is prompt text, not a claim of certification or affiliation.

Pro-informed VidBee extensions

Reference checkout: VideoLingoPro
Revision: dcb707f10978b0f088b1c334985155b531f32c8a
Inspected files (SHA-256):
  core/prompts.py: 97252e2f8cb16ca0805e37c08cd37a5bc7823c41ebf9ae2bf3f1c45780ca1d67
  core/_5_split_sub.py: e80528b3ee42a9670e73d4ba37f1a57b6149ceae0fc9d7ed40350473c8c27d82
  core/text_guard.py: 0963491254cbbd740c09b5630c30086b6d0858aae45879ab18b5c0c56c0b4690

The following are VidBee-authored instruction changes informed by observed Pro behavior, not a copy of the Pro application or an assertion about its license:
- align.txt compares align1/align2, validates the chosen candidate and tries the other candidate before failing. Both candidates retain the existing indexed source/target schema. Punctuation-only target parts are rejected.
- subtitle-constraints.txt reconstructs each source cue and checks exact text before saving. Unlike Pro's diagnostic-only normalized-text guard, the Agent must reject a changed source candidate; this remains an instruction-driven check, not a new backend validator.
- number-to-text.txt is an additional VidBee-authored prompt for same-language, context-sensitive numeric pronunciation, preserving identifiers, units and significant digits. It uses the converted field and returns sparse speechOverrides bound to exact saved cues.
- create_dubbed_video consumes separate spoken variants, retains display subtitles and source timing, and supports bounded meaning-preserving spoken repairs after real fitting errors. It does not adopt Pro's fuzzy time mapping or unbounded effective tempo. Actual cue binding and audio fit are enforced by Desktop; language quality remains the Agent's responsibility.
