Refined Rules for Whisper Agent Enhancements | Generated by AI
Rule rewritten. Key adjustments vs. the original:
- Comprehensive detail is now the core directive — explicit “do NOT summarize,” length-similar-to-input, examples of forbidden hand-waves like “various topics” / “and so on,” and a list of specifics to preserve (function names, file paths, version numbers, model names).
- Numeric normalization ported from
transcript_agent.pyand expanded: “fifty K CNY” → “50K CNY”, “GPT four” → “GPT-4”, “two point five times” → “2.5x”. - Whisper-specific speaker rule (vs. transcript_agent’s A:/B: labels): no labels at all, third-person narration, with explicit Q&A phrasing guidance.
- Tone preservation now calls out humor / frustration / hesitation explicitly.
- Reorganization: explicit permission to break original chronological order and group scattered topics together.
- Language: always English; translate-while-refining for Chinese with canonical-form rule for brand/technical names (e.g., “CUDA” not “kuda”).
- Output format: hard rule — start directly with H1, no preamble.
Re-run the same command to see the difference:
ww whisper refine /Users/lzwjava/projects/private-projects/whisper-playground/yangyang/yangyang.txt
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋