zmm-cut

zmm-cut: talking-head finished-video editing

It handles "after the shoot." Topic selection, scriptwriting, and review before the shoot belong to /zmm-topic /zmm-script /zmm-review。

🔴 Boundary: only delete and reorder, never add words (clarified 2026-09-05)

This only handlesthe version he actually said. Within that version:

✅ Can do ❌ Cannot do

Delete (dead shots, duplicates, filler words, jargon, cut-off endings) Addany word he did not say

Reorder(adjust order at the container layer, bring the result to the front) Smooth sentences, change word order, replace wording

Merge pickup shots into the original footage Reorder the segment where he derives the conclusion (see references/内容层重组.md §II.3)

⚠️ "This skill does not change the copy" means it does not add words, not that nothing may be touched. Deleting and reordering according to copywriting rulesis exactly this skill's main battlefield。

✅ Can do	❌ Cannot do
Delete (dead shots, duplicates, filler words, jargon, cut-off endings)	Addany word he did not say
Reorder(adjust order at the container layer, bring the result to the front)	Smooth sentences, change word order, replace wording
Merge pickup shots into the original footage	Reorder the segment where he derives the conclusion (see `references/内容层重组.md` §II.3)

0. Startup order (do not skip)

🔴 Probe the environment first, then decide how to proceed(corrected 2026-09-05):
- A host with the ChatCut plugin-version prerequisite skill → load it first,do not bundle it in the same turn as ChatCut tool calls
- Desktop MCP direct connection(tool names look like chatcut_desktop/*）→ Without that prerequisite skill, per server requirements first run a read-only locate (get_active_project) to confirm the target project
- None of them present / reports signed out → use references/ChatCut实操.md §I'sdegraded path: Tell them to download + local transcription bypass(all content-layer work does not require ChatCut)
chatcut-talking-head-guide — execution spec for talking-head editing
Read zmm/references/家族公约.md(if it cannot be read → say plainly "the convention cannot be read, the red lines cannot be guaranteed" and stop)
- zmm/references/交互规范.md（🔴 not just read once and done: at the end follow §IV's three-piece set — Recap · Before/After · next step given as numbered options; if information is missing, per §IV usemultiple choicequestions,ask only one at a time; for cases where it does not apply see §V)
Read {vault}/00-规则与索引/口播拍摄基线.md(time/place/outfit fixed once and for all)
Read memory {vault}/08-技能记忆/zmm-cut/ + _通用/
🔴 When content must be changed (not just removing verbal tics), read references/内容层重组.md
🔴 Before calling a ChatCut tool for the first time, read references/ChatCut实操.md(three tool pitfalls encountered in real testing)
🔴 After editing you must return to 口播稿输出格式.md §IV's "upload version" — see §V of this document

This skill's built-in criteria are in references/规则卡.md(criteria / why / how to check / strength), read it once before starting work;{vault} when there is a corresponding rules file in it, the vault takes precedence, with the rules card as the fallback.

⚠️ Read on demand, do not read it all at once. This skill only covers the "shot → final cut" stage, you do not need to read the topic-selection layer or scriptwriting layer files.

1. Account default parameters (🔴 locked in, do not ask the user again)

🟢 Set by 詹明明 on 2026-08-27.These are default values, not questions to confirm every time. If the user has not said to change them, just use them; if the user says otherwise, follow what the user says this time.

⚠️ The table below shows measured values for "this one account", not general recommendations. If there is config , thenfrom config read; if it cannot be read, use the table below, andstate that example values are being used. When switching accounts / switching topics, the main workflow still runs as usual,only this table changes。

Item	Default value
Platform	Douyin (vertical 9:16)
Final cut duration	Keep the original length —— do not cut content just to make it short (the criterion is absolute watch time, not completion rate)
Overall speed	🔴 1.15×(after voice editingafterwardapply uniform speed-up, not segment-by-segment adjustment)
Default processing	① Voice editing(remove verbal tics, pauses, repetitions) ② Subtitles
Not done by default	B-roll · MG animation · background music —— only do it if the user explicitly asks
Rhythm style	Tight and punchy

Subtitle specs (🔴 locked in)

Item	Value
Font	HarmonyOS Sans Bold
Stroke	Slight
Shadow	Yes, light
Overall look	Clean and modern
Highlight	★ Yes, color yellow `#FFE20A`
🔴 Who chooses the highlighted keywords	The user chooses them. Provide candidates when delivering the draft,do not decide for them

2. Workflow (seven steps, must stop between each step)

🔴 talking-head-guide Hard rule:There are dependencies between multiple processing stages, so the final cut must be locked in order, and each major step must be confirmed with the user separately; do not package multiple checkpoints into one reply. If the upstream is changed, everything downstream must be redone (subtitles written against the pre-speed-up timeline = all wasted).

⓪ 素材清点与合并 → 停，说清有几条素材、哪条是补拍、要不要合并
① 素材进项目     → 停，确认转录出来了
② 语音剪辑       → 停，确认「他实际说的」这一版对不对
③ 整体加速       → 停，确认节奏（倍速见 §一）
④ 字幕 + 高亮     → 停，让用户选高亮关键词
⑤ 交付           → 只有用户明确说要导出/下载才导
⑥ 导出后验文件本身 → 抽帧确认字幕/B-roll/打码都烧进去了

⓪ Material inventory and merging (added 2026-09-05)

The skill originally assumed there was only one piece of material. In practice it is often two or more(original footage + pickup shots, multiple cameras, multiple recordings).

Firsttranscribe each one, read them separately,do not splice them right away
State clearlywhich one is the main material and which one is the pickup, and where the pickup shouldcover
🔴 Once a pickup is added, you must do a "verbatim repetition scan" — see references/内容层重组.md §II.2
Mergein the new order, not by appending the pickup at the end

① Footage enters the project

First list_projects / create_project / target_project lock in the project,then do everything else
Once the project is createdimmediatelybring up the editor (preview_start), so the user can watch the progress
Footage can come in three ways: the user uploads directly in the editor · a local path is given in the session (via asset-import) · the user has already uploaded it (first browse_assets look for it,don't just say you can't find it right away）
You can start working as soon as the transcript is out, no need to wait for the raw bytes to finish uploading (A-roll only needs the text transcript)

② Voice editing (the main battlefield of this skill)

🔴 This step has two layers, don't mix them together: Mechanical layer(fixed verbal tics, batch pauses) doesn't require understanding the meaning, tools can do it; Content layer(what to keep, what to delete, what to move)must understand the meaning, the criteria are in references/内容层重组.md。

⚠️ Doing only the mechanical layer = just wiping the film clean, not improving it. When the user says "cut it", they usually want both layers,but every deletion or change in the content layer must have a clear reason。

Strictly follow the talking-head-guide A-roll process:

read_script → read through timeline.md, see the structure clearly
First run the mechanical cleanup（clean_script): only handle fixed verbal tics (uh/um/hmm) and long pauses
🔴 After running it, you must re-read the refreshed timeline.md — what you read before is already outdated
Then docontent layerjudgment — follow the references/内容层重组.md five criteria: identify junk shots · verbatim repetition scan(mandatory for pickups) · reasoning chains must not be rearranged for suspense · the opening must have both a hook and resonance · before deleting, ask "is it still clear after deleting this"
apply_script lands on the timeline
Read it back to check:seam logic(is a bridge missing) · consistency before and after · does the relative time still hold · is the ending complete · are the pauses too tight

🔴 Three things specific to this account, don't cut them out by general rules

Keep	Why
Colloquial noise(ne / a / ha / reduplication / slip-of-the-tongue repetition)	Rule every sentence being neat = the audience hears in a second that it's being read from a script. ⚠️ The general "remove verbal tics" will clear these out too — keeping one every 5–8 lines is by design, not a mistake
Inserted sentences inviting self-reflection("think about it carefully", "I don't know if you've ever had a moment like this")	Rule it carries an action instruction, not a filler word
Qualifiers(mostly / a lot of the time / maybe that's what)	Rule must not be changed to absolutes

🔴 Criterion: after this word is deleted,Is this sentence cleaner now, or does it now read like a script? Reads like a script = don't delete.

Pause handling

Obvious long pauses (>0.8–1s) → compress to about 0.3s
Keep 0.3–0.5s between sentences — don't compress it away
Keep pauses at turns, contrasts, and emphasis, those are by design (where the script is marked ⏸⏸)
Normal breathing within a sentence → leave untouched

③ Overall speed-up 1.15×

🔴 Must be done after the voice edit is finalized, if the order is reversed all subtitles have to be redone. Uniform speed-up,not segment-by-segment speed adjustment — adjusting segment by segment makes the tone speed up and slow down unpredictably.

④ Subtitles + highlights

Follow the talking-head-guide captions spec, apply the fixed spec in §I.

🔴 I'm not allowed to decide the highlighted keywords for him. Method:

Read through the subtitles and pick outone every 8–12 secondscandidate keywords (pick judgment sentences, numbers, transition words, core words in punchline sentences)
List them for the user to choose, clearly mark at which second and what the original sentence is
Only add highlights after the user checks them off

⚠️ Don't highlight whole sentences — the purpose of highlighting is to make the eye pause while scrolling; highlighting a whole sentence is the same as no highlighting.

⑤ Delivery

🔴 The default delivery is an "editable ChatCut timeline", not an MP4. Only when the user explicitly saysexport / render / download / final deliverydo you use submit_export → track_export。 "cut it a bit" "clean it up" "make a version" don't count as export intent.

IIb, 🔴 Acceptance discipline: metadata lies (added 2026-09-05, tested three times)

Line count, text, and image can only be judged by full-size rendered frames.

Fooled by	Actual
Subtitle API reported "one line, no overflow"	Rendered as two lines
In the multi-grid thumbnail it looked like the text was wrong	A full-size single frame proved it was misread
Tool receipt reported "cleaned 1 word"	The timeline was actually 4.9 seconds shorter

After every major step, always verify two things：

Numbers — frame count / duration / segment count, do they match expectations
Image — renderfull-sizeframes,don't draw conclusions from thumbnails

⚠️ Thumbnails are only for rough screening of "which frame to pick". See references/ChatCut实操.md §III.

III. Red lines (unique to the editing stage)

	Rule
Don't change content	The script was finalized at shoot time. Editing only deletes, never adds,don't fill in words, smooth sentences, or change word order for him
Image information points	Outfits / desk setup / background / wall decorations are all information points.When editing, don't cut out frames that carry information
No AI screenshots	In the segment about AI, the imageshould not be paired with AI interface screenshots — it narrows the audience (`不要变成工具号`）
Position for the big text	In the script's "points that need handling," the lines marked with ⏸ for big text,must be matched during editing
Blur amounts	Any frame showing backend data/payments/customer information must be blurred

IV. Who it's for

One person doing their own shooting and editing. Shooting, editing, and publishing are all done by him alone:

The plan must be executable by one person alone; do not assume there is an editor
Give judgments and actions directly; do not write "for reference"
Zero jargon: don't say "A-roll," say "the main track where he's talking"; don't say "ripple," say "the ones after it will shift forward too"

V. 🔴 Must be backfilled after editing (this step is the easiest to miss)

After the final cut is published, transcribe the actually broadcast text back and write it into the "upload version" section of the script. For the spec, see {vault}/00-规则与索引/口播稿输出格式.md "upload version" under §IV.

Three parts, none may be missing:

Full upload version text(code block + line numbers + breath points) — 🔴 only fix proper nouns and typos that ASR misheard,do not change word order, do not smooth out sentences
Difference table(script version vs what he actually said) — use only five types: add / delete / reword / restructure / reorder
Rules distilled from the differences → add to the knowledge base

🔴 Why it cannot be skipped: the script is written, the broadcast version is what he said. Every difference between the two is a vote he cast with his mouth — that is the most honest feedback in this system.

⚠️ Exception: in the difference table, the "reword" columndoes not apply "default to his being better"—— speaking is more prone to on-the-spot dulling than writing.Wherever the AI version is sharper, flag it separately and ask.

VI. Absolutely do not

Do not call ChatCut tools before the environment has been fully probed (see §0.1)
Do not use apply_script(preview:true) as a read-only preview — it will error while alsopartially committing
Do not reuse old Card ids after changing subtitle font size/box width — it will repaginate; the id is a content hash
Do not use tool receipts and metadata as the basis for acceptance — see §IIb
Do not make up ChatCut's features, paths, prices, or version numbers — product facts that change must be checked against the current page of the official Docs
Do not package multiple checkpoints into one reply
Do not add background music / B-roll / MG / transitions on your own initiative — if the user didn't ask, don't do it
Do not choose highlight keywords for the user
Do not use local ffmpeg to compress a flattened MP4 as the main deliverable (ffmpeg is only for read-only inspection of source files)
Do not export just because of "a quick edit"

VII. Memory

Self-check before finishing:

Which editing approach did the user reject (record as "correction," e.g., "colloquial noise must not be cleaned")?
Which parameter was empirically verified as effective (record as "effective method," with values)?
Which pitfall of the tool was hit(record as "correction,"with forensic methods: how was it discovered, how was the wrong guess falsified)?

Write to {vault}/08-技能记忆/zmm-cut/，check for duplicates first。

Don't know the next step → return to /zmm。