Aicwos AI Voiceover Copywriting Companion

Problem

Many voiceover scripts fail not because of weak ideas, but because the context is fragmented: lecturer style lives in chat notes, product docs are long files, previous series episodes require rereading full text, and banned or required phrases are easy to miss. Aicwos organizes this material into a local database so script generation loads only the necessary fragments instead of stuffing whole files into context.

How It Works

• Lecturer profiles: lecturer_analyzer.py builds a style profile from sample scripts; later runs load a lite profile with --action lite, while management uses the full profile.
• Knowledge retrieval: SQLite + FTS5 + vector search returns relevant paragraphs; without the semantic model, it falls back to keyword search.
• Series continuity: --action summaries reads prior episode summaries rather than full transcripts, reducing repeated context.
• Behavior rules: allowed and prohibited rules are managed separately, with allowed rules taking priority; the generated context includes these rules automatically.
• Data boundary: files are the edit layer, the database is the query layer; private knowledge overrides public knowledge, and deletes move to a trash bin first.

Limits

It is best for short-video voiceover scripts, lecturer style management, and paragraph-level knowledge retrieval, not general-purpose writing. First use requires initializing a local “控制台” directory and database; cloud knowledge sync is optional and not required for local search. Script length should stay within a 10% deviation, and rule conflicts should be handled through the intended layers instead of bypassing the file layer to edit the database directly.