Talking-head speech cleanup

This is the A-roll speech-cleanup playbook for Pireel Studio. It is the SAME content the read_editing_guide MCP tool returns — with this skill loaded you do not need to call that tool. This is judgment, not new tools: execute with the tools you already have.

How cuts map to tools

• Main narration: read_script gives sentences in SOURCE-video seconds. Remove spoken passages with cut_narration (pass those source-second ranges; it converts to the edited timeline, compresses overlays, and re-lays captions).
• Inserted [clip X] segments: their own clock — cut with cut_range (edited seconds) or drop the whole segment with delete_shot.
• Explain edits to the user by the actual spoken words, NEVER by segment ids or timestamps — the user can't see those.

Workflow (any "clean up / tighten / de-um the narration" request = run this end to end)

1. Orient — read_script if the transcript isn't already in the conversation. Read the whole thing; understand the goal and the content structure, and reconstruct full sentences across ASR rows.
2. Decide the removals by the rules below — for a long transcript work section by section. Collect EVERY source-second range to drop: hesitation fillers, failed retakes, false starts, abandoned fragments, dead pauses that belong to a removed attempt. Keep complete units; when unsure, don't add the range. FIRST look at the very first and very last rows — the recording's pre-roll and post-roll (see Head & tail); mid-body cleanup alone leaves that junk in.
3. Apply them in ONE cut_narration call (ranges = the whole list). Do NOT cut one range per call.
4. Review — re-read what the viewer will now hear (broken logic, missing context, over/under-cut, wrong order); fix only clear problems, then tell the user what you cleaned up in plain words.

For plain cleanup / de-umming, just run it. For aggressive shortening, restructuring, highlight/short-version, or a generated hook, confirm target length + structure + what to preserve with the user BEFORE step 3.