The core of cheat-seed isdiscussing topic selection with the user, not mechanically brainstorming. Good content comes from the user's real experiences + observations + emotions — these are things the AI cannot possibly brainstorm out of thin air. The AI's role islisten to the user → help distill angles → write a draft, not dump 15 candidates for the user to pick from.
Default mode: conversational, one at a time.
Batch mode(--batch N): keep the old brainstorm N candidates + write N drafts process, for users who "have no ideas at all + want to initialize in bulk".
Three Modes (auto-detected)
Mode A — 用户主动给主题(**最常见**):
用户:"/cheat-seed" + 直接说"我想做一条关于 X 的"
或:"/cheat-seed 我最近开会被领导..."
↓
AI 围绕 X / 这件事**深挖**——什么瞬间触发?最让你 [情绪 / 不爽 / 觉得有意思] 的是哪点?
↓
收敛到一个具体角度 → 提议 → 用户认可 → 写 1 份 draft → 完成
↓
问"下一篇?" 或用户说"今天就这样"
Mode B — the user gives a direction but not specifics: User: "Recently I want to do something about [workplace / marriage and relationships / AI / ...]" ↓ AI: "[Scope] is too broad. What specific thing you've recently encountered made you want to do this direction?" ↓ Converge to Mode A's specific experience
Mode C — the user has no ideas at all (rare): User: "I don't know what to do" / "Help me think of a topic" ↓ AI: "Okay, entering brainstorm mode — first grab trending topics + your previous interest directions, and give you 1 suggestion" ↓ Run trend-sources to grab trending topics + read candidates.md / predictions/ to see the user's history ↓ Propose 1 angle (not 5) → user approves → write draft
Batch Mode — the user explicitly asks for batch (/cheat-seed --batch 5):
Follow the old brainstorm process: 3 questions → 15 candidates → user picks → write 5 drafts.
For users who "want to get the next 2 weeks of topics done all at once today".
**关键纠正**(与旧版的区别):
- AI **不主动开放问**——等用户给输入再深挖
- 一次一个选题,不是 5 个
- 默认对话式 + 一次一个,batch 是 escape hatch
## Constants
- **DEFAULT_TREND_SOURCES = ["manual-paste"]** — 仅 Mode C / Mode A 灰色场景 / Batch 用到。用户可在 state 里加 aihot / trendradar-mcp
- **TREND_TOOL_ROUTING** — 按 `content_form` 路由数据源,详见 [shared-references/data-source-routing.md](../../shared-references/data-source-routing.md)
- **MODE_B_MAX_REPROBE_TURNS = 2** — Mode B "为什么" 反问最多 2 轮;超过则转 Mode C
- **MAX_DEEP_DIVE_TURNS = 4** — Mode A 收敛阶段最多 4 轮反问,避免 AI 过度盘问
- **WITH_DRAFT = yes** — 默认确认角度后立刻写 draft;用户可说 "等下,我自己写" 跳过
- **DRAFT_LENGTH** — 派生自 `state.typical_duration_seconds`:30s→100-200字 / 90s→250-500字 / 240s→600-1000字 / 450s→1100-2000字 / 900s→2200+字
- **HUMANIZE_DRAFT = on**(默认)/ off —— 写完 draft 后用 `humanizer` skill 过一遍,去掉 AI 写作 tells(em-dash 滥用 / rule of three / inflated 词汇 / 空泛归因等)。off 时直接出原始 AI draft。**只 humanize 正文,不动 header 的"必须改写"警告**
## Inputs
| 必填 | 来源 |
|---|---|
| `.cheat-state.json` | 读 calibration_samples / typical_duration / cadence |
| `rubric_notes.md` | 读当前 rubric(粗打分用) |
| `script_patterns.md` | 读已有 pattern(写 draft 时按 cheat sheet 选结构) |
| `predictions/*.md`(如有) | 已发历史,brainstorm 时作为 context |
| `audience.md`(如有) | 受众画像——选题 / 写稿时的"谁在看"镜子(由 `/cheat-persona` 派生) |
## Workflow
### Phase 0: 前置检查 + 加载所有 context(**核心:3 个 context 来源**)
1. 读 `.cheat-state.json` → 不存在则提示先跑 `/cheat-init`
2. 读 `rubric_notes.md` 拿当前公式(粗打分用)
3. 读 `script_patterns.md`——写 draft 时按 cheat sheet 选结构
4. **读已有 prediction 文件**(含 init 时 import 的 reconstructed)作为 **context 来源 A**(用户自己历史):
- 0 个 → A 为空
- ≥1 个 → A 有内容,提取 (title / 7 维 / 实绩)
5. **读 `benchmark.md`**(如存在)作为 **context 来源 B**(对标账号):
- `state.benchmark_status = imported` → B 有内容,提取对标账号的样本主题分布、调性、Pattern
- `state.benchmark_status = none / pending` → B 为空
6. **读 `audience.md`**(如存在且非空骨架)作为 **受众镜子**:
- Confidence 🟡 以上(有真实复盘数据派生)→ 选题 / 写 draft 时作为"这个 persona 会在乎吗"的检验视角
- Confidence 🔴/🟠(空 / 仅 benchmark seed)→ 当弱参考,不当硬约束
- **不进粗打分**——audience.md 是 creative lens,粗打分仍只用 rubric。persona 影响"写什么 / 哪个角度",不影响分数
7. 检查用户的入参——是否含具体 topic / 经历,决定走 Mode A/B/C/Batch
**brainstorm 时的 context 优先级**(**Claude 判断**——下面是参考默认):
- **A 主导**(用户自己数据):当 Claude 判断用户数据已能驱动方向时(参考默认:`calibration_samples ≥ 10`,但 Claude 可以更早——如 N=5 但出现 ≥3 个与 benchmark 明显不一致的强样本)
- **B 主导**(benchmark):用户数据少 + benchmark 有内容时
- **B 缺席**(benchmark 为空)+ 用户数据少:纯靠用户 input + 抓热点;明确告诉用户"没 benchmark 也没足够自己数据,建议跑 /cheat-learn-from 后再回来 brainstorm"
判断依据**不是死磕样本数**,而是看:
- 用户最近 N 个样本的实绩**是否与 benchmark 的高表现样本类型一致**——一致说明 benchmark 仍有借鉴价值;不一致说明用户已经走自己的路
- 用户的样本**多样性**——3 篇都是同类内容不算成熟;3 篇覆盖不同类目反而比 10 篇同类更可信
### Phase 1: Mode 分流
读用户输入,识别:
**含具体名词 + 情绪 / 经历词**("我昨天开会..." / "我看到 X 让我..." / "我对 Y 觉得...")→ **Mode A**(深挖;如话题是时事,Phase 2A.5 询问要不要拉外部数据)
**含方向词但无具体内容**("想做职场" / "AI 方向" / "婚恋")→ **Mode B**(单问"为什么想做这个"——用户内省窗口,**不调任何热点工具**)
**显式说没想法**("不知道做什么" / "帮我想" / "随便给个")→ **Mode C**(按 [data-source-routing.md](../../shared-references/data-source-routing.md) 调热点工具 + 用户挑 + 回到内省;都不行则走"聊经历"三选项兜底)
**显式 `--batch N`**(用户主动批量)→ **Batch Mode**
**纯 `/cheat-seed` 无附加内容** → **询问入口问题**:
What do you want to do today?
- Have a topic / experience you want to do → tell me directly ("I want to make an X" / "I recently X...")
- Know the general direction → tell me ("want to do workplace" / "AI direction") → I'll ask you individually why you want to do it
- No ideas at all → say "help me think" → I'll use [aihot / trendradar] to pull today's trending topics for you to see
- Get it done in bulk → say "batch "
(I won't chase you with a bunch of open questions — give me one sentence and I'll start)
注意这是**唯一的开放式问题**——只在用户**纯触发** `/cheat-seed` 时问。如果用户已经在触发词里给了内容("/cheat-seed 我想做..." 或 "找选题 我最近开会..."),直接进 Mode A/B/C 不再问这一句。
### Phase 2A: Mode A 深挖(用户给了具体 topic / 经历)
**核心原则**:围绕用户给的内容**深挖**,**不要切到别的话题**。
**反问类型(按场景挑)**:
- 触发瞬间:"你说 X 这件事,最初是哪个具体瞬间触发你想做的?" / "是什么让你觉得这值得讲一条视频?"
- 情绪锚点:"这里面最让你 [生气 / 觉得荒唐 / 觉得有意思] 的是哪个细节?"
- 角度选择:"你想说的是 [角度 a:现象批判] 还是 [角度 b:自我反思] 还是 [角度 c:泛化到普遍]?"
- 受众想象:"你心里想着是说给哪种人听?她/他听完会怎么想 / 怎么转发?"
- 反对意见探测:"如果有人反驳说 [反方观点 X],你会怎么回?"——逼用户先想清楚立场
**反问纪律**:
- 一次只问 **1 个**问题(不要塞 3 个连珠炮)
- 最多 `MAX_DEEP_DIVE_TURNS` 轮(默认 4)——超过就主动收敛:"OK 我感觉够了,帮你提议一个角度试试"
- 用户的回答如果含 emoji / 简短 / 不耐烦 → 立刻收敛,不要逼
**收敛输出**:
I feel this angle could work:
[One-sentence thesis: within 50 characters]
Approach:
- Use [Pattern X structure] (from script_patterns.md)
- Hook: [specific scene / sentence]
- Body: [what the 3 observations are]
- Ending: [MVP sentence direction]
Rough scoring (v0 equal-weight 7 dimensions): ER=X HP=X QL=X NA=X AB=X SR=X SAT=X → composite ≈ X.X Confidence: 🔴 Extremely low (you've only calibrated 0/N pieces)
Do you want me to write a draft first? (yes / change angle / I'll write it myself)
用户回 yes → Phase 4 写 draft。
用户说"换角度" → 回 Phase 2A 深挖更多。
用户说"我自己写" → 把 candidate 加进 candidates.md 标 tier1,结束。
### Phase 2B: Mode B — 单问"为什么",触发用户内省
用户给了方向但不具体("想做职场" / "AI 方向" / "婚恋")。**这阶段不调任何热点工具**——是用户内省的窗口,外部信息会污染。
只问一个问题,**单刀直入**:
Why do you want to do this topic?
不要问"a/b/c 三种例子"——那是 dump 选项让用户挑,破坏内省。让用户自己组织语言。
**根据用户回答分流**:
| 用户答 | 分类 | 行为 |
|---|---|---|
| 含具体经历 / 个人卡点("我自己经常加班" / "我看到 X 让我...") | **真动机** | 转 Mode A 深挖(Phase 2A) |
| 抽象热度归因("这话题最近热" / "别人都在做" / "听说能涨粉") | **空动机** | 反问"那你自己对这个话题最有感觉的角度是啥?"——逼出个人 stake;继续空 → 转 Mode C |
| "我也不知道" / "朋友说赚钱" / 模糊推搡 | **真没想法** | 直接转 Mode C |
**反问纪律**:最多 2 轮。第 2 轮还问不出真动机 → 直接转 Mode C,**不要无限挖**。
> 设计意图:Mode B 是"过滤器",不是"工厂"。用户来这里要么暴露真实动机(→ 进 Mode A)要么暴露自己其实没想法(→ 进 Mode C)。两条都比硬要在 Mode B 里产出选题更好。
### Phase 2C: Mode C — 外部素材 + 用户挑 + 回到内省
用户完全没想法(直接显式说 / 从 Mode B 转过来)。**这是唯一默认调热点工具的入口**。
按 [shared-references/data-source-routing.md](../../shared-references/data-source-routing.md) 的路由规则:
1. **拉外部素材**(按 `content_form` 选 trend source):
- `tutorial-builder` / AI 类 → 调 aihot skill
- `opinion-video` / `long-essay` / `podcast` / `other` → 调 trendradar-mcp(如启用)
- `mixed` → 两个都调
- 都不可用 → 走 manual-paste(询问用户:"今天看到啥可以拍的?粘几条 URL/标题")
2. **聊经历兜底**(用户拒绝看外部素材 / 外部素材都不感兴趣):
You don't feel anything about the external material, so come back to yourself. Three openings, pick one to start talking about:
a) A specific thing you've actually encountered recently? ("Last week I saw my colleague X...") b) Something you've recently read/seen that made you want to complain? ("There's an answer on Zhihu...") c) Some unsolved confusion you've been mulling over for a long time? ("I've never understood why X...")
Just pick one and start talking.
3. **拉到外部素材后**,用 rubric 粗筛 + 按 content_form 过滤,留 5 条最契合的:
These 5 fit your shape today:
- [Title A] (Source: trendradar / Weibo Hot Search / hot_score: 8.5)
- [Title B] (Source: aihot / Model category / Featured)
- ...
Which one resonates most with you? If none do, reply 'none', and I'll switch directions and ask.
4. **用户挑了一条 → 回到内省**:
OK [Title X]. Why do you resonate most with this one? Is it [angle1] or [angle2] or something else?
→ 用户答 → 转 Mode A 深挖。
5. **用户回 "都没"** → 转回 Mode C 第 2 步的"聊经历兜底"。
**关键**:热点不是"塞 5 候选让用户挑",是"给材料 + 强制问用户的个人 stake"——AI 不替用户决定哪条最值得做。
### Phase 2A.5: Mode A 灰色场景 — 用户讲了时事话题
Mode A 默认深挖用户经历。但如果**用户讲的本身是时事话题**(产品名 + 时间词 / 人名 + 事件词),按 [data-source-routing.md](../../shared-references/data-source-routing.md) 的"时事判定"规则,**询问**用户要不要拉外部数据作参考:
💡 [Topic] is current affairs—I can pull today's public opinion trends (sentiment across platforms + main angles) for reference.
Want to see it? Reply 'see' and I'll pull it; reply 'no need' and I'll dig straight into your angle with you.
用户回 "看" → 调对应 trend source → 把数据 inline 到深挖 context;
用户回 "不用" → 标准 Mode A 深挖,不动外部数据。
**永远不主动调**——用户的 angle 优先于外部数据,避免外部信息**带偏**用户视角。
### Phase 2D: Batch Mode(用户显式 `--batch N`)
**保留旧 brainstorm 流程**:
1. 问 3 个清单问题(兴趣 / 调性 / 红线)—— Batch 模式才问这些
2. 抓热点 + Claude brainstorm 15 候选
3. 用户挑 N
4. 写 N 份 draft 到 scripts/——**每份都走 Phase 4 的段落版格式 + Phase 4.5 自检**(line-format + humanizer),不因为批量就跳过
详见 commit history(旧 cheat-seed 的 Phase 1-3)。这是 escape hatch,不是默认。
### Phase 3: 计算 candidate ID + 落候选池
不管 Mode A/B/C 哪条路径,确认角度后:
1. 算 candidate id:`sha256("seed-" + 立意 + 触发时间)[:12]`
2. 写一行 entry 到 `candidates.md`(按 [candidate-schema.md](../../shared-references/candidate-schema.md) 格式)
3. 标 `tier=tier1` + `read_status=deep_read`(已经讨论过,不是 skim)
### Phase 4: 写 draft
`WITH_DRAFT=yes` → 顺次写到 `scripts/<YYYY-MM-DD>_<id>_<short-title>.md`:
**写 draft 前必读** `script_patterns.md` —— 按"结构选型 cheat sheet"对应用户的 topic 选合适结构。如果文件还在抽象骨架阶段(用户没填几个 pattern),就用 starter rubric 对应的通用框架。
**字数**:按 `DRAFT_LENGTH` 派生(基于 `typical_duration_seconds`)。
#### ⚠️ 正文必须是段落版,不是字幕格式(**最常见的生成跑偏**)
模型的训练先验会把"视频脚本"默认写成提词器/字幕的短行格式。**这是错的**——cheat-seed 的 draft 是给用户**改写**的散文稿,不是拍摄字幕。字幕是剪映拍后自动断的,不是写作时的形态。
生成正文时,眼睛盯住下面这个对照:
❌ Subtitle format (don't write like this): Have you noticed All reviewers are saying the same thing Your research is too outdated But if you look closely Everything they cite is from 5 years ago reactions
✅ Paragraph version (you must write like this): Have you noticed that all reviewers are saying the same thing—your research is too outdated. But if you look closely, everything they cite is from 5 years ago reactions. AI isn't new; what's new is that this time everyone has collectively awakened.
规则:
- **每段 100-300 字**,逗号 / 句号 / 破折号自然连,**不在句子边界硬断行**
- 段与段之间空一行(自然的主题切换才换段)
- 一份 draft 正文一般 3-6 段,**不该有几十个单句行**
#### 格式:
```markdown
# [立意标题]
> ⚠️ **Draft by Claude — 你必须改写后再拍**
>
> 这是脚手架,不是成品。你的语气 / 节奏 / 个人经历无法 AI 生成。
> 改写流程:
> 1. **直接在本文件改写**(同 path:scripts/<...>.md)
> - 加你的语气、个人经历、真实金句
> - 砍铺垫、砍模型缩写、砍学术包装
> 2. 改完后跑 `/cheat-predict scripts/<本文件>.md`
> 3. 拍完跑 `/cheat-shoot scripts/<本文件>.md`
**Article ID**: <12 位 hash>
**调性**: [基于讨论得出的,不是清单 Q]
**目标时长**: <state.typical_duration_seconds 转换> 分钟
**目标字数**: <按时长派生>
**结构选型**: [按 script_patterns.md 的 cheat sheet 显式标,如 "metaphor 优先" / "数据反转开场"]
**用到的 patterns**: [编号 + 简短说明]
**讨论种子**: [一句话回顾 deep dive 出来的核心]
---
[draft 正文 —— **段落版**,3-6 段,每段 100-300 字,不是单句碎行]
WITH_DRAFT=no(User says "I'll write it myself") → skip Phase 4 + Phase 4.5.
Phase 4.5: draft self-check pass (format + remove AI flavor)
After the draft is written and saved to disk,before showing it to the userrun two self-check steps.Fixed order: first 4.5a fix format, then 4.5b remove AI flavor—humanizer processes prose; feeding it the fragmented lines of subtitle format will make a mess.
Why it's safe(doesn't pollute calibration): the cheat-seed draft is not what gets predicted/published—after the user rewrites it, what cheat-predict scores isthe user's final draft. These two steps just give the user a cleaner starting point.
Phase 4.5a: line-format self-check (subtitle format → paragraph version)
Phase 4's prose instructions + ❌/✅ comparison already suppress the prior at generation time, but generation can still go off track. This step isa deterministic fallback:
# 只看正文段(--- 分隔线之后)
body=$(awk '/^---$/{f=1;next} f' scripts/<id>.md)
line_count=$(printf '%s\n' "$body" | grep -c .) # 非空行数
char_count=$(printf '%s' "$body" | wc -m | tr -d ' ')
avg_chars_per_line=$(( char_count / (line_count > 0 ? line_count : 1) ))
Criteria:avg_chars_per_line < 15 and line_count >= 8 → judged as subtitle format → auto-reflow:
- merge hard line breaks at sentence boundaries back into natural paragraphs
- split into 3-6 paragraphs by topic shifts, each 100-300 characters
- use Edit to replace the body paragraphs (header unchanged)
- in the Phase 5 output mark a line: "📐 Subtitle format detected, reflowed into paragraph version"
no match → skip, the body is already in paragraph version.
Phase 4.5b: humanizer remove AI flavor
HUMANIZE_DRAFT=on(default)—use humanizer skill to run through it once. Claude's own first drafts naturally carry AI tells (em-dash overuse / rule of three / "inflated" vocabulary / vague attributions / -ing shallow analysis); this step clears them out.
Steps:
- Check whether
humanizerskill is available (~/.claude/skills/humanizer/exists):- not available → skip 4.5b, add a line to the Phase 5 output "(humanizer not installed, draft is the original AI version—
git clone https://github.com/blader/humanizerto ~/.claude/skills/ to enable automatic AI-flavor removal)"
- not available → skip 4.5b, add a line to the Phase 5 output "(humanizer not installed, draft is the original AI version—
- Available → call via the Skill tool
humanizer,pass only the draft body(---after the separator, the already-reflowed paragraph version from 4.5a),never pass the header:- the header's
⚠️ Draft by Claude — 你必须改写后再拍warning isan intentional scaffolding marker, not prose to be humanized - voice calibration: if the user has historical scripts (
videos/*/script.md) or has filled inscript_patterns.md, pass the most recent 1-2 as voice reference samples for the humanizer—have it lean toward "this user's voice" rather than "generic human voice"
- the header's
- the humanizer returns de-AI-ified body text → use Edit to replace the body paragraphs of the draft file (header untouched)
- record the "which tells were fixed" reported by the humanizer (e.g.
em-dash 滥用 ×3 / rule of three ×2 / inflated 词汇: "深刻" "本质上"), show it in the Phase 5 output
Discipline:
- the humanizer isde-AI-ifying, notrewriting on behalf of the user. It makes the draft less machine-written, butit is still not the user's voice—the header's "must rewrite" warning still holds, and the Phase 5 output must reiterate it
- if the humanizer changes a sentence so it deviates from
结构选型/ the pattern used → the pattern prevails, roll back that sentence (the pattern was decided in discussion with the user; the humanizer should not overturn it) - humanizer not responsible for layout—line-break issues were already fixed in 4.5a, and what the humanizer receives is already the paragraph version
Phase 5: output "next steps" + ask whether to continue
✅ Draft 写完:scripts/2026-05-04_<id>_<short>.md
📐 版式自检:通过(段落版) ← 或"检测到字幕格式,已重排为段落版"
🧹 humanizer 过了一遍:修了 em-dash 滥用 ×3 / rule of three ×2 / inflated 词汇 2 处
(draft 现在不那么"机器味"了——但这仍是脚手架,不是你的声音)
接下来你可以:
- 改写这份 draft(直接在原文件改)—— 加你的语气、经历、真实金句
- 改完跑 "打分这篇 scripts/<...>.md" 看 7 维评分
- 决定要拍 → "启动预测 scripts/<...>.md"
下一篇你想做什么?
(直接告诉我具体经历 / topic,或者说"今天就这样"结束)
the humanizer line appears only when
HUMANIZE_DRAFT=onand the skill is available. When not installed, replace it with a line on how to enable it.
User says "that's it for today" → end cheat-seed. User gives a new topic → go back to Phase 1 and re-triage.
Key Rules
- AI does not proactively ask open questions—only when the user purely triggers
/cheat-seeddoes it ask the entry question once; otherwisewait for the user to provide input before digging deeper - one topic at a time—by default Modes A/B/C each give 1 suggestion; only go to Batch if the user actively asks for a batch
- Questioning discipline: ask 1 at a time, at most 4 rounds; if the user gets impatient, converge immediately
- Dig deeper around the topic the user gave, don't switch to something else—if you say "got chewed out by my boss in a meeting," the AI should not ask "so have you recently felt that AI makes everyone..." that kind of parallel topic
- Writing a draft must read script_patterns.md—choose the structure according to the user's existing patterns
- the draft is scaffolding—add a prominent warning "must rewrite" in the header
- the humanizer only de-AI-ifies, it does not rewrite on behalf of the user—Phase 4.5b makes the draft less machine-like, but it is still not the user's voice; the "must rewrite" warning does not become invalid just because it went through the humanizer
- the body is a paragraph version, not subtitle format—when generating, watch the ❌/✅ comparison in Phase 4; Phase 4.5a uses
avg_chars_per_line < 15 且行数 ≥ 8as a deterministic fallback, and reflows on a hit. Subtitles are auto-broken by Jianying after shooting, not the form at writing time
Refusals
- "Skip the deep dive, write a draft directly" → ask "Do you want to give me a topic and have me write it directly? OK, but the draft quality may be poor—I don't know your angle. Give me a one-sentence thesis and I'll write it"
- "AI decides the topic for me" → refuse. In the Mode A/B/C paths, AI always onlypresents external material + asks the user for their angle, and does not make the final call for the user on "which one to do"
- "For Mode B I can't be bothered to answer why, just give me 5 candidates" → refuse. The "why" in Mode B is a filter—if you can't answer it, you shouldn't use Mode B's direction. Either go to Mode A and give specific experiences, or go to Mode C and I'll help you find material
- "In Mode A, just pull trending topics for me directly, don't ask for my consent" → refuse. In Mode A the user already has an angle; pulling external data without permission would contaminate their perspective. See data-source-routing.md
- "Write 5 drafts at once" → not in the default flow. The user must explicitly
--batch 5 - "I can't be bothered to rewrite, just post the AI draft directly" → warn "Posts generated directly by AI have lower ER when posted and will contaminate your calibration data", but if the user insists, it is also allowed (mark
unmodified_ai_draft: true)
Integration
- Upstream:
/cheat-initAt the end of Phase 5, whenpool_status=none + calibration_samples=0, proactively ask "Run /cheat-seed now?" - Upstream:
/cheat-recommendWhen candidates is empty, mention/cheat-seed - Upstream:
/cheat-statusWhenpool_status=none + 距 init >24h, prompt "Not posted yet—run /cheat-seed?" - Downstream: the user's candidate → candidates.md (tier1, already deep_read)
- Downstream: (default) draft → Phase 4.5 humanizer to remove AI flavor → scripts/.md → user rewrite → /cheat-predict
- Optional dependency:
humanizerskill (MIT, external project). When installed in~/.claude/skills/humanizer/, Phase 4.5 is automatically enabled; if not installed, it is gracefully skipped.Not bundled into cheat-on-content—the user clones it themselves - Difference from
/cheat-trends: cheat-seed isdiscussion + writing drafts(conversation-heavy); cheat-trends ismulti-adapter scraping + rough scoring(fetch-heavy). The two have different purposes and do not replace each other.