talking-head-cleanup

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)

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.
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.
Apply them in ONE cut_narration call (ranges = the whole list). Do NOT cut one range per call.
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.

Core principle

Remove defects without changing meaning; make speech smoother, not harder. Prefer small local cuts over whole-sentence deletion. Good cleanup keeps the logic coherent, the expression clearer, the delivery natural, and preserves the speaker's intent, tone, and rhythm. When unsure whether a cut harms meaning, logic, or flow — keep it, or make a smaller cut.

Edit by complete semantic units

Move / keep / delete complete sentences, ideas, answers, or steps. Do not cut half a sentence just because a few words match.
read_script rows are ASR segments, NOT semantic units: one sentence / idea / retake may span several rows, and one row may hold only part of a sentence. Reconstruct the full spoken idea across adjacent rows BEFORE choosing the cut boundary.

Fillers — two tiers

Tier 1 (safe to remove when they carry no meaning): um, uh, er, ah, 呃, 额.
Tier 2 (decide by ROLE, never by word list alone): so, like, then, 然后, 就是, 那个, 那, 对, 所以, 但是, 嗯, 啊. Keep it when it carries sequence, continuation, contrast, cause, reference, response, emphasis, or natural tone; remove only pure hesitation / padding. If removing it makes the splice sound hard, keep it. If unsure, keep it.

Retakes (the speaker retries the same idea)

Treat as a retake ONLY when multiple attempts say the SAME intended idea — not intentional repetition / emphasis, and not a second pass that adds new info or tone.
Keep ONE complete, natural version; cut only the failed or fully-covered part. The cut starts at the repeated / failed idea, not automatically at the earlier setup or transition.
Prefer the later attempt (usually closest to intent) — but if it is missing needed context / subject / conclusion, keep the more complete earlier version or preserve the earlier lead-in.
Keep one natural lead-in / section marker; drop only the extra restarts. NEVER stitch unfinished fragments from different attempts into one artificial sentence.

False starts / unfinished fragments

Remove only when the fragment clearly forms no useful info: an abandoned thought with a complete version later, a dangling phrase ("this is actually…") with no completion, or the leftover start of a failed attempt.
Keep an imperfect sentence that still carries useful info, and keep a lead-in that supplies the subject / object / context needed later. If only part of a sentence is bad, cut the smallest bad span; if a local cut cannot sound natural, keep the whole thing.

Head & tail — the recording's pre-roll / post-roll

Easy to miss: explicitly scan the very first and very last rows.

Head, before the real opening line: drop countdowns ("3, 2, 1"), "okay / let me start / let's go / 开始 / 来 / 那我开始了", throat-clears, mic/test phrases, and dead air, plus any abandoned false-start intro that a later take replaces. Start the video on the first real content word.
Tail, after the real ending: drop "okay / that's it / that's a wrap / cut / 好了 / 就这样 / 停 / 完了吧", "did I get that / was that good / 刚才那条行吗", trailing dead air, and the reach to stop recording. End on the last real content word.
But KEEP a genuine sign-off / CTA / outro ("thanks for watching, subscribe", 记得点赞关注) — that is content, not logistics. Cut recording-logistics chatter, never a real closing.

Preserve connective tissue

List labels, contrast words, subjects, verbs, and adjacent words are NOT filler when removing them makes a kept idea ungrammatical, abrupt, or misleading. Trim the smallest span that keeps the line speakable.

Pauses (scope note)

Pause / silence micro-compression is not a primitive here — cleanup means removing spoken passages by the transcript. Do NOT create a black gap on the main track as a pacing "pause" (no source playing = black frame). If a beat needs air, leave the source silence or cover it with a graphic; only cut a pause when the failed attempt around it is being removed with it.

Tasks beyond cleanup (same principles)

Highlight / short version / restructure / target-script: when the task NAMES what to keep, trim to that boundary — start and end at the requested words and drop the off-script head / tail of the segment they sit in; do not over-keep a whole segment for one requested sentence. When reordering, move complete units and re-check that connectors ("so / but / next / 这 / 所以") still work.
Confirm first for aggressive shortening, structural changes, or a generated hook: align target length, structure direction, and what to preserve before editing.

After cutting

Re-read what the viewer will now hear: broken logic, missing context, over-deletion, wrong order, delivery too rushed. Fix only clear problems. Captions re-lay automatically from the new timeline — don't hand-fix them for a cut.