What problem does it solve? LLM agents that walk a transcript document by hand and write rows themselves silently fail: they narrate an ingest that never happened, mangle long Drive ids, and report success with zero rows written. This Skill removes that failure by delegating all reading, parsing, and writing to a deterministic ingest tool, leaving the agent to judge dictionary corrections and report verified results. ## Core Features & Use Cases - Deterministic transcript ingest: Runs the scribe-ingest tool against a Google Doc transcript, writing conversations, participants, and transcript_lines into BigQuery with resumable cursors and verified outcomes. - Folder planning: Runs the tool in plan mode to list which documents in a client's Drive folder still need ingesting, resuming, or re-processing after edits. - Dictionary-based corrections: Scores correction candidates against active project_dictionary terms and applies approved fixes to transcript_lines, while never touching human-made edits in line_edits. - Use Case: Claire asks what transcripts remain for a client; you run the plan command, dispatch ingests per document, apply high-confidence corrections, and report outcomes, warnings, and human-edit conflicts verbatim from the tool's JSON. ## Quick Start Ask the agent to ingest the named interview transcript from the client's Drive folder into BigQuery and report the outcome, line count, and any warnings.
