Phase 1 Before locking the 3rd item, theory, read zmm/references/理论框架选型与装配.md §1–4, first compare explanations that are genuinely different, then score, verify, and confirm according to the existing "lock the four things". §5 mapping is merged into the one-pager, with no separate approval; the concept must still provide transferable judgment and cannot replace definition with philosophical labels.
First read zmm/references/家族公约.md(if unreadable → say plainly "the convention cannot be read, the red line cannot be guaranteed" and stop,do not substitute memory for the convention），
This skill's built-in criteria are in references/规则卡.md(criteria / why / how to check / strength), read them once before starting;{vault} when there is a corresponding rules file, the vault prevails and the rule card is the base.
Then read zmm/references/交互规范.md（🔴 it is not reading once and counting it as done: for wrap-up follow §4, the three-piece set — Recap · Before/After · next step give numbered options; when information is missing, follow §4 and usemultiple-choice questionsto ask,ask only one at a time; for cases where it does not apply, see §5), then read {vault}/00-规则与索引/文案框架-总入口.md(the only navigation).
🔴 Its position on the shortest path:replaces /zmm-script, not added beside it
选题 → 【重讲概念型走这儿】→ 红线审核 → 发
/zmm-concept 代替 /zmm-script
⚠️ zmm the router remembers one measured fact: this system once had 7 pre-publication gates, and the result was 13 finished drafts, 0 published.
The number of gates itself is the cause of not publishing.
🔴 This skill adds no new gates. It does only one thing:move rework that would already happen forward so it happens on one page.
Why this skill exists (measured cost, do not delete)
N22 "The lower the price, the worse it sells" went through 12 versions. Of the 12 versions, only two really mattered: ① Zhan Mingming said "my experience cannot be the main body, Douyin's mass audience cannot reach it」 ② he spoke for himself。
The other nine versions were guessing what he wanted, and the reason each guess was wrong was the same: the full draft was written first, and only then was it discovered that the proposition level was wrong.
| Version | Where it died |
|---|---|
| v1 · v2 · v4 | The ontology definition was not locked, and "his experience" was used to prop it up |
| v3 | The theory was not locked(Weber was explaining "it dropped but there was no feeling", not "it dropped and nobody bought") |
| v5 · v6 | The proposition was not locked(it changed from "tell one point" to "tell the whole thing") |
⇒ Six versions of rework, all dying on the same four things. And each of those four things needs only one sentence.
🟢 N23 "There is no precise traffic" verified that it works — but pushed the problem downstream (2026-08-29)
N23's retelling v1 was basically usable, and it only iterated to v4, with not a single proposition-level rework. Locking the four things holds.
🔴 But everything this one lost was lost in Phase 3–5:
| What was lost | Which step it died at | Cost |
|---|---|---|
| The sentence spoken on behalf of the audience(The sole source of the impulse to forward) | Phase 4 — I wrote it into "suggest adding it back," he approved it,I only landed the other spot, this sentence never made it into the body text | It's not in the final cut. This item's forwarding criterioncan't be read out |
| The landing point changed and it wasn't treated as a script change | After Phase 3 — capability knife → classification knife, and the landing-point sentence is "the script's ID card" | frontmatter / title / filename once still carried the old landing point |
| Rejected the right opening based on the wrong pacing parameter | Phase 2 — used the ÷5.1 point value to calculate the first 5 seconds' budget, judged that it "doesn't fit with high-ticket" | He added it back in the final cut,it fits. Measured 6.88, the parameter was off by 35% |
⇒ Once the upstream is locked, rework doesn't disappear — it moves downstream.
The four spots in this update (Phase 2 pacing · Phase 3 post-check landing point · Phase 4 landing in body text · Phase 5 cover and writing to disk) all come from N23, all are downstream.
Who it's spoken to
A solo knowledge creator who self-sources and self-edits. Three hard constraints: don't assume he has a team · no upward reporting (give the judgment directly) · zero jargon(the theory still applies, the terms just don't appear in the final script).
🔴 The cross-cutting layer is always there：CLAUDE.md Floor item 7 — After finishing each paragraph, reread it once using "I," and write out the inner line he'd have at that moment. If you can't write it, this paragraph isn't finished yet. No hint of blame or preaching is allowed anywhere in the piece.
Process
Phase 0 接题 + 两道准入闸门
Phase 1 🔴 锁四样（一页纸，不写整稿） ← 本技能的全部价值
Phase 2 写整稿（一次写完，不逐段确认）
Phase 3 🔴 交给他口述（D1）
Phase 4 只剪不扩 + 红线全表 + 差异表 + 规则入库
Phase 5 /zmm-title 归口 + 封面 + 落盘 + 建取数日历
Phase 0 · Take the topic + admission (no pointless questions)
The topic comes from {config.paths.topic_pipeline} the topic pool in its directory, or is newly given by the user.
Two gates, if either is missing, don't do it(the criteria are in the pool's §2.3, go read them, don't go by memory):
| # | Gate | One sentence |
|---|---|---|
| ① | Anchor point | Before he hears your new formulation,is there a word he was already familiar with that you first took apart？ |
| ② | Ruler criterion | The thing you hand over,can he take it and measure something you never mentioned? If it can't measure, it's still just a definition |
By the way, run it once through the deduplication gate（总入口 §2 layer 1): core conclusion × main analogy comparison + write out which item is the sibling。
Phase 1 · 🔴 Lock four things (the core of this skill, see references/锁四样.md）
Ask it all in one go, one page, don't write the whole script.
| # | What to lock | Who gives it |
|---|---|---|
| 1 | Specific question | AI proposes candidates,he decides |
| 2 | 🔴 Ontology definition / the counterintuitive line | only he can give it |
| 3 | Classic theory + disciplinary common-knowledge target | AI finds it and scores it on five dimensions,he confirms |
| 4 | 2–3 plain-language analogies | AI proposes,he swaps |
🔴 If the four things aren't all locked, you're not allowed to write the whole script.
For the theory in item 3, each one gets a card; any that lacks the line "which level it's been verified to" is not allowed into the final script:
| Row | What to write |
|---|---|
| What problem does it solve | One sentence |
| Mechanism | One sentence of cause and effect |
| Source | Name · Work · Year · Chapter or page (write it down to whatever level you can reach) |
| To what level was it verified | 读过原文 / 权威二手（教材、综述，写明是哪本） / 没核到 |
| Verbatim or paraphrase | The sentence quoted in the final draft: is it a verbatim translation or a condensed paraphrase; a paraphrase must be marked in the draft with "the gist is" |
Rules: If the same source cannot be verified twice in a row, write 没核到 and stop; do not expand the search further, and do not name it in the final draft; for a quotation seen through secondary literature, write "cited in" and do not pretend to have read the original; when a personal name first appears, write "Country · Role · Chinese name (common foreign name)".Naming a book in the final draft is equivalent to handing the audience a verifiable source; the audience really will go check. "I remember he said" and "I have read the original" must look different on this card.
⚠️ Item 2 is the easiest to skip, and it is precisely the root of six rounds of rework. AI cannot write sentences like "pricing itself is part of the product" — it can only write functional descriptions ("pricing answers three questions"). If you cannot get item 2, stop here and wait; do not substitute theory or personal experience.
Phase 2 · Write the full draft (write it all in one go)
Follow {vault}/04-模板/概念纠偏型口播稿模板.md six sections.
Within section ⑤, follow zmm-script/references/概念型口播.md seven beats —— 🔴 must be opened and read; the two beats are the easiest to miss：
② Say the audience's answer for them("the easiest beat to miss, determines success or failure") · ⑥ Second cross-industry scenario("the entire sense of gain comes from this beat").
When the draft is out, report three things:duration range · gap distribution table(calculate seconds) · red-line full-table scan result。
🔴 Give only a duration range, not a point value（2026-08-29 总入口 §3 #2 re-cut):
原速区间 = 稿面汉字 ÷ 6.9 ~ ÷ 5.1
🔴🔴 2026-08-31 correction: the basis for the passage above was wrong; do not follow the old version. Actual measurement confirms the final video used 1.2× speed ⇒ 6.88 is the final-video speed, not the speaking speed。 Original speed = 1583 ÷ (230×1.2) = 5.74 characters/second.
⇒ The reason written on 08-29 when overturning the old ruling was "N19 5.1 vs N23 6.88,35% dispersion」—— after removing the speed multiplier it is 5.1 vs 5.74, 12.5% dispersion. That reason no longer holds. 🔴 The three conceptual original-speed speaking rates are actually quite concentrated：N19 5.1 · N20 5.32 · N23 5.74. ⇒ Closed:calculate original-speed speaking rate using the 5.3–5.8 characters/second range(multiple actual measurements); for the first 5 seconds, use the lower bound of the range to calculate capacity and the upper bound to calculate duration; final-video figures are accelerated numbers, so divide by the speed multiplier before putting them into the parameter table.
🔴 Draft output action (write the four frontmatter fields together; if one is missing it does not count as filled back)：
稿面字数 / 成片秒数 / 倍速系数 / 原速语速 = 字数 ÷ (成片秒数 × 倍速)
⚠️ Do not swap the two rates: useoriginal speedto calculate script length; usefinal-video speed
(Douyin's 2s 跳出 / 5s 完播 / 观看趋势 are all on the final-video timeline).
🔴🔴 Before using duration to reject a sentence, ask one question: would a 35% error flip this conclusion? N23 actual measurement cost: Counting at ÷5.1, only 21 characters fit in the first 5 seconds, and on that basis I rejected the "and high customer spend" he wanted to add. In his final cut it was added back, and it fit. —— When a parameter is uncertain, don't use it as a veto.
🔴 When handing it over, you must state plainly: this version is a draft for you to narrate from, not a final cut.
Phase 3 · 🔴 Hand it over for him to narrate
No retelling, no shooting. Narrated back → archived as-is → then processed.
🔴 First thing after the narration comes back: check whether the landing point changed (added in N23)
It's not the wording you check, it's the landing point. 口播稿输出格式：The landing-point sentence is this script's ID card — swapping the landing-point sentence means swapping the script, not revising it.
Ask one question: the landing point he said, and the one you locked in —are they the same sentence?
N23 measured: no, and his is better.
| AI script | His retelling | |
|---|---|---|
| Landing point | When you're not running paid traffic, you have no "aiming" action | You can have precision marketing, but you shouldn't have precision traffic |
| What kind of knife | Capability knife(you can't do it) | ★ Classification knife(that thing goes by another name) |
⇒ Landing point changed = change four things on the spot: frontmatter landing-point sentence · work title · cover big text · file name。 🔴 Don't drag it to right before publishing — N23 once still had the old landing point hanging on it.
⚠️ A changed landing point usually means it got better(he's reasoning, the AI is describing).This step is to capture it, not to correct him.
Phase 4 · Cut only, don't expand
Optimization can only cut from the retold script; you can't expand upward from any AI script.
Four must-dos：
- Full red-line table（
违禁词避雷表10 categories + absolutes + symbol level) — 🔴 It's not grepping four words - Unify person(when explaining a mechanism, use a generic term consistently throughout; only switch when describing a specific industry scenario, and switch only in that passage)
- Internal consistency（
/zmm-reviewquestion 6 (skeleton and breakpoints): do all numbers/times/facts in the text line up front to back) - Difference table + rules into the library（
总入口§5 four steps: archive the exact words → judge the attribution layer → write it as an executable criterion → attach the measured cost)
🔴🔴 The fifth (added in N23, because I missed it):Anything approved to be added back must be put into the body on the spot and verified with grep
Because of the rule "can't expand upward from an AI script," the things in the difference table that "he deleted"can only be listed and asked about, not added back on your own. But after he approves? — the original process broke off right here.
N23 measured cost: the sentence spoken on the audience's behalf (theonlysource of the share impulse), after he approved adding it back, I only landed another spot; this sentence never entered the body ⇒ it's not in the final cut ⇒ that criterion can't be read out this time.
🔴 Three steps, none may be missing：
① 他批准 → ② 当场落进正文（不是写进"建议"栏） → ③ grep 原句验证在不在
⚠️ The same thing has already been taught once: 📣 A checklist quoting the original sentence but not the line number, becauseline numbers will silently point to a different sentence。 This one is of the same kind: a sentence written in the suggestions columnwill silently not exist。 Test: after the change, grep that original sentence. If grep finds nothing = it isn't there.
Phase 5 · Title cover + visuals + saving to disk
Title coverGoes through /zmm-title as its owning channel(must mark structural numbers S1–S12, and must answer "based on this title, does the opening of the body need changing").
Final draft goes to {vault}/_草稿/NN_认知_{标题}.md, format per 00-规则与索引/口播稿输出格式.md。
The filename and the draft's destination must match — if the destination changes, rename it; don't drag it out until right before publishing.
🔴 Where are the cover assets (added in N23: I got the directory wrong once)
docs/brand/design-system.md ← 视觉系统（暖白纸底 · 档案绿 · 宋体黑 · 刻度线母题）
docs/brand/douyin/cover-template.html ← 抖音封面模板，改帧和标题即可
docs/brand/douyin/cover-00X.* ← 已发的封面，看它们对齐风格
⚠️ Not in _装修/。 In N23 I only searched _装修 and _已发布，didn't search docs/,
and ended up making a version of a white-background black-text pure-text cover, detached from the whole visual system, scrapped and redone.
The template's two hard rules(written in the template comments, go read them):
- Cover large text ≤6 characters per line, at most two lines（
口播稿输出格式2026-08-29 wording corrected: originally wrote "≤6 characters", missing "per line") - When both lines are 6 characters, the font size must be reduced(long-sentence mode 108/148), otherwise the size contrast is gone and the template falls apart
🔴🔴 The pitfall of grabbing cover frames: subtitles are burned into the visuals
The template's title area lands right on the subtitle band. N23 measured: the subtitle band is at y1400–1520 (75% of the frame).
Don't just grab a frame at random. Sample the whole video at 1.5-second intervals → programmatically filter for clean subtitle bands → then sort by contrast in the head area. N23 measured: of 143 frames, only 4 had no subtitles.
🔴 After publishing (added in N23)
| What to do | Why |
|---|---|
Register _已发布/已发布内容库.md — core conclusion + main analogyFill in both keys |
Missing one means you can't judge duplication next time. ⚠️ N20 was published and never registered; the dedup gate is blind to it |
| Create a T+24h / T+7d data-pull calendar | T+24h is a snapshot, the whole account pulls in the same window; reading late makes cross-comparison impossible |
| Hard-code into the calendar descriptionthe data-pull list + minimum play-count threshold | Falling below the threshold = don't read it this time,not that it read out a negative result |
| Copy "what this one can't be attributed to" into the T+7d entry | To prevent skipping over it during review |
🔴🔴 Open the next draft's frontmatter 承接上条复盘: column, and copy item by item this one's review's "N things that land in N+1", each marked ✅ 已执行 / ❌ 已否决（理由） |
Measured: N20's review fixed "remove the 1.15× speedup for the next one" —N23 not only didn't remove it, it raised it to 1.2×. The instruction wasn't rejected, it was silently lost.Cost: the editing layer and the copy layer have never once been separated to this day; that 24pp retention collapse in N23 could have been a clean reading.Pre-publish gate: this field is not allowed to have blank entries. Vetoing is fine, silently losing is not. |
🔴 Six "must nots" (all have empirically tested costs; not one may be softened)
| # | Must not | Cost |
|---|---|---|
| 1 | Write the full draft before all four are locked | N22 v1–v6 Six rounds of rework |
| 2 | Use "my experience" as the skeleton | Three criteria: ① If you delete the whole "I" passage, does the argument still hold? ② If it takes up >1/4 of the body text, it is the main subject ③ 🔴 Does the Douyin general audience have this scenario? If not = it can only serve as a single piece of supporting evidenceN22 passed only after being cut from 31% to 5% |
| 3 | Only grep the four prohibited words | 「USD" mixed all the way into the example of 作者方法论 §5.2 — a prohibited word that gets into a long-term file will replicate itself in every subsequent session |
| 4 | Cite rules from memory | I cited 语言风格档案 one item but had never read that file (that time I guessed right).The §6 loading list says that to write the body text you must read it, so you really have to read it |
| 5 | Reconcile conflicting rules yourself | Check 总入口 §3 ruling table;list anything not in the table and ask, then after the ruling write it back into the table(this skill has already produced #13/#14/#15) |
| 6 | He makes the call to keep it, deletes it, then adds it back himself | 裁定 #12: execute, and write the exemption next to it |
Load-on-demand checklist (saves tokens; do not read it all at once)
| What I am doing now | Read at minimum these |
|---|---|
| Judge whether this concept can be done | This skill's Phase 0 + topic pool + 已发布内容库 |
| Lock the four | references/锁四样.md(the only must-read file included with this skill) |
| Write the body text | 04-模板/概念纠偏型口播稿模板.md + zmm-script/references/概念型口播.md + 语言风格档案 |
| Pair with theory | Three admission criteria + five-dimension scoring + theory card "to which level it is verified" (general disciplinary knowledge does not follow citation discipline） |
| Red-line self-check | 规则总表 condensed checklist + 违禁词避雷表 full table |
| Review | /zmm-review four decisions and eleven questions (when the vault has 发布前审核流程.md , use that as authoritative) |
Absolutely never do
- Do not write the full draft before all four are locked.
- Do not draw the conclusion "this is a misconception" for the audience — you only expose the crack; that line is for them to say themselves.
- Do not do any conversion action — no "comment 1 / DM / link in the comments / contact me if you need it" may appear.
- Do not use examples the audience cannot reach (for luxury examples, look atwhether the audience stands in the seller's or buyer's position）。
- Do not force in a book just to have an endorsement (
§1.65if mechanism match <2, eliminate directly). - Do not add new gates.
Don't know the next step → return to /zmm。