Whisper Refine Prompt Expansion | Generated by AI

Home 2026.05

File parses cleanly. Here’s a summary of what changed:

File: /Users/lzwjava/projects/ww/ww/audio/whisper_refine.py

What changed: The REFINE_RULE prompt string expanded from ~1,200 words to ~3,900 words — roughly 3x more comprehensive.

What was added (10 numbered sections, each expanded):

  1. Anti-summarization preface — Strong upfront warning that LLMs naturally condense, with explicit “stop and write out details” directive and a “200 paragraphs in, 200 paragraphs out” rule.

  2. Content & Structure — Title: specific-not-generic rule with examples. Summary: must touch every H2 section. H2/H3 guidance: don’t force H3 when a single topic. Sequence preservation: don’t regroup chronological walkthroughs. Topic grouping across time: merge scattered discussion with “circled back” markers.

  3. Detail Preservation (the biggest expansion):
    • Anti-summarization rules with a CRITICAL TEST (scan input vs output for omissions)
    • Collapse-only-same-speaker-redundancy exception
    • Expanded “NOT to do” examples (5 concrete before/after pairs)
    • Numbers/dates: 10+ conversion examples (percentages, milliseconds, quarters, GB, currency)
    • Approximation markers: “roughly 100” vs guessing
    • Ambiguous numbers: “(15 or 50?)” not guessing
    • Anti-pattern list: 13 phrases to never write (“etc.”, “among other things”, “several approaches”, etc.)
    • Code formatting: inline code and fenced code blocks with language specifiers
    • Architecture: descriptive prose, no ASCII art
  4. Action Items, Decisions, Commitments (entirely new section):
    • Decision: format, conditional decisions, explicit “not to” decisions, rationale required
    • Action item: owner + task + deadline format, unassigned items, concrete-verifiable rule
    • Action items grouped per-section under H3, not in a global appendix
    • Open questions, deferred topics, no-consensus markers

  5. Tone & Fidelity — Expanded with: concrete before/after examples showing scrubbed vs preserved tone. Metaphor preservation. Emphasis handling (bold/italics). Intensity gradient (strong agreement → leaning). Disagreement and contradiction rules — present both sides fairly, note walk-backs, don’t synthesize false consensus.

  6. Speaker Handling — Added: do NOT infer roles from content. Do NOT consistently associate positions with voices. Expanded third-person phrase bank (15 patterns). Direct quotation rules (2-3 per article max, no quotes for paraphrasing). Back-and-forth grouping pattern.

  7. Noise Removal — Expanded filler list (“er”, “ah”, “mm”, “so”, “well”, “I think” as hedge). Added “keep” section with examples of substantive “yeah”/”exactly”. Added test: if removal changes meaning, keep it.

  8. Reorganization — Added: tangent handling (off-topic cat → omit; related technical tangent → its own section). Transitional phrase guidance (one per major shift, not between sub-topics).

  9. Language — Added: code-switching section with Chinese-English mixed examples. Pinyin + parenthetical for untranslatable concepts. Company name canonization table (ByteDance, Tencent, Alibaba). Terminology consistency rule.

  10. Output Format — Added: strict first-character rule. No trailing whitespace/separators. Markdown conventions (ATX headings, - lists, bold/italic/blockquote/table rules). Final quality check — 8-item mental checklist before finalizing.

  11. Technical Formatting (entirely new section): Inline code rules for 13 categories (function names, files, commands, env vars, ports, etc.). Fenced code block rules with language specifiers. Architecture description rules.
Back Donate