Recording Transcript Structured Note Generation

Problem

Transcripts from meetings, interviews, and trainings are often fragmented, colloquial, and inconsistent, with homophone errors, repetitions, and missing speaker labels. Teams that copy raw text into documents risk presenting guesses as facts or filling in names, numbers, owners, deadlines, and conclusions that were never stated. The core constraint here is evidence-bound editing: keep the result readable without inventing missing information.

How It Works

• Output level: choose concise, standard, or detailed mode; a vague clean-up request defaults to standard.
• Fact separation: separate confirmed statements, inferences, and transcription gaps, using markers such as 【原文未明确】 when the source does not specify a fact.
• Speaker mapping: build confirmed, possible, or unknown speaker identities from context clues, not voice recognition.
• Type-aware structuring: detect meeting, training, interview, or work-log content and apply the matching template; long inputs are processed with segmented summarization, evidence IDs, and deduplication.
• Word delivery: generate structured content as JSON, validate it through the build pipeline, and check layout issues instead of dumping raw text into a document.

Boundaries

This skill fits reviewable, actionable note preparation rather than verbatim transcript production, real-time transcription, or speaker identification by voice. Missing critical facts are surfaced as gaps rather than filled in automatically, and Word output claims remain limited to what the environment can actually verify.