Use when a user provides a call, meeting, interview, podcast, or other conversation transcript that needs to be rewritten into clean speaker-by-speaker dialogue, especially when the transcript is long, noisy, speaker labels are messy, or ASR errors need conservative repair.
Turn noisy transcript text into readable dialogue without adding meaning that is not present in the source.
Inputs
Require the raw transcript text.
Expect a speaker roster before the transcript, with each speaker described in 2-3 words.
Treat the speaker roster as primary attribution context.
Use the roster heavily to resolve vague labels, role references, and likely speaker continuity.
Reuse raw speaker labels only when they help map a turn to the roster.
If the roster is missing or incomplete, stay conservative: do not invent names.
Workflow
Use a controller workflow with sub-agents so the main context holds conversation structure, not every raw utterance. The controller must not load the full raw transcript into its own context.
Read the speaker roster and the user's task framing only. Do not load the full raw transcript into the controller context.
Dispatch a mapper sub-agent to read the full raw transcript and return a compact conversation_map with:
likely speakers and source labels
topic shifts and natural chunk boundaries
相关技能
ambiguity hotspots such as cross-talk, broken sentences, or unclear terms
Create and maintain a continuity_ledger containing canonical speaker names, stable technical terms, unresolved mappings, and any other consistency decisions.
Split the transcript at natural boundaries with a small overlap. Short transcripts can stay as one chunk.
Dispatch a cleaner sub-agent for each chunk with the chunk text, the relevant conversation_map excerpt, the speaker roster, and the current continuity_ledger. Each cleaner should return:
cleaned_chunk: rewritten dialogue for that span
open_questions: anything still ambiguous after best-effort cleanup
Only when a chunk contains a real ambiguity hotspot, dispatch a challenger sub-agent for that span and ask for 2 conservative hotspot_variants. Use this for uncertainty, not for stylistic experimentation.
The controller chooses the most defensible variant, updates the continuity_ledger, harmonizes overlaps and naming, and assembles the final transcript.
When seams, speaker continuity, or hotspot choices remain uncertain, dispatch a follow-up sub-agent to inspect only the relevant raw span and return a minimal excerpt or resolution note. The controller should not reopen the raw transcript itself.
Controller Rules
Do not paste, quote, or load the entire raw transcript into the controller context.
Keep the controller working set limited to the speaker roster, conversation_map, continuity_ledger, rewritten chunks, and small targeted raw excerpts returned by sub-agents.
If more raw evidence is needed, ask a sub-agent to inspect the specific span and return only the minimum needed to decide.
Use sub-agents for transcript inspection. Use the controller for coordination, consistency, and final assembly.
Cleaner Rules
Keep speaker order and conversational flow intact.
Format each turn as Speaker: utterance or Speaker [uncertain]: utterance.
Remove silence hallucinations and filler loops.
Remove most repeated yeah or similar filler insertions unless they answer a yes/no question, confirm something explicitly, or carry real meaning.
Remove filler, false starts, stutters, and repetitions only when meaning does not change.
Fix proper nouns, numbers, contractions, grammar, agreement, and technical terms only when nearby transcript context, repeated usage, or the speaker roster strongly supports the repair.
Rewrite fragments only enough to recover the intended meaning.
Preserve all concepts, claims, decisions, numbers, and technical details present in the source.
Do not add new information, interpretation, speaker intent, or implied context that is not grounded in the transcript.
Do not summarize, combine turns, or smooth wording in ways that introduce new concepts.
When a word or segment cannot be recovered confidently after best-effort cleanup, write [unclear].
Use Speaker: utterance when the transcript and roster support the attribution confidently.
Use Speaker [uncertain]: utterance when a named speaker is plausible but not certain.
If no named attribution is defensible, preserve the source label and mark it [uncertain] when that helps, otherwise use Unknown Speaker [uncertain]: utterance.
Output
Return only the rewritten transcript inside <rewritten_transcript> tags.
Internal workflow artifacts such as conversation_map, continuity_ledger, cleaned_chunk, and hotspot_variants are for coordination only. Do not expose them in the final answer.
If you made any significant interpretation, non-obvious speaker mapping, or material repair that could affect how the transcript is read, add a <notes> block after the transcript. Omit <notes> when there is nothing material to flag.
Use this exact shape:
<rewritten_transcript>
Alice: First cleaned utterance.
Bob [uncertain]: Second cleaned utterance.
Unknown Speaker [uncertain]: Third cleaned utterance.
</rewritten_transcript>
<notes>
- Brief explanation of a major repair or non-obvious speaker mapping, if needed.
</notes>